{
  "source": "https://agentsandboxes.org",
  "license": "CC BY 4.0",
  "reviewed": "2026-11-18",
  "disclaimer": "General information, provided as is and without warranty. Not professional advice. Product details change; confirm them with each provider.",
  "terms": "https://agentsandboxes.org/terms/",
  "providers": [
    {
      "id": "e2b",
      "name": "E2B",
      "company": "E2B",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://e2b.dev/about.md"
      },
      "kind": "hosted",
      "tier": 1,
      "alternatives": true,
      "website": "https://e2b.dev",
      "docs": "https://docs.e2b.dev",
      "pricing_url": "https://e2b.dev/pricing",
      "github": "https://github.com/e2b-dev/E2B",
      "tagline": "Open-source cloud sandboxes for AI agents, each a Firecracker microVM, with pause/resume, snapshots and forking.",
      "isolation": "Firecracker microVM with its own kernel per sandbox",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": true,
      "sdks": [
        "Python",
        "JavaScript/TypeScript"
      ],
      "cold_start": null,
      "max_session": "Up to 1 hour continuous (Hobby), 24 hours (Pro) or custom (Enterprise); pausing resets the clock and paused sandboxes are kept until deleted",
      "max_session_short": "1 h / 24 h (Pro)",
      "persistence": "Pause/resume preserves filesystem and memory; snapshots of running sandboxes, forking from a snapshot, filesystem-only snapshots, auto-resume and volumes (beta)",
      "snapshots": true,
      "pause_resume": true,
      "gpu": false,
      "gpu_detail": "No — sandboxes are CPU-only",
      "networking": "Internet on by default; can be disabled or restricted with allow/deny lists of IPs, CIDRs and domains; public sandbox URLs for exposed ports; custom domains and published egress IPs",
      "pricing": "Per second: $0.000014 per vCPU-second (~$0.050/vCPU-hour) + $0.0000045 per GiB-second (~$0.016/GiB-hour). Pro plan $150/month plus usage; Enterprise from $3,000/month",
      "free_tier": "Hobby plan: $100 one-time credit, 10 concurrent sandboxes, 1-hour sessions",
      "compliance": "SOC 2 Type II; HIPAA BAA and DPA available on request",
      "regions": "Default region plus a shared EU cluster on request",
      "best_for": "Code interpreters and short-to-medium-lived agent tasks that need microVM isolation, with an open-source self-host path",
      "overview": [
        "E2B is the best-known agent sandbox and the default many AI frameworks reach for. Every sandbox is a Firecracker microVM with its own kernel, created in a call from the Python or JavaScript SDK. Beyond basic command execution, E2B ships a **Code Interpreter SDK** (Jupyter-style stateful execution with charts) and a **Desktop SDK** for computer-use agents.",
        "State features have matured: pause/resume preserves memory and filesystem, running sandboxes can be snapshotted and forked, and paused sandboxes are kept indefinitely. The SDKs and backend runtime are Apache-2.0, with a self-hostable single-node package (“E2B Embed”) and Enterprise BYOC on AWS, Google Cloud or Azure."
      ],
      "notable": [
        "Raised a $21M Series A in July 2025, led by Insight Partners.",
        "Backend runtime (e2b-dev/runtime) is open source and self-hostable as “E2B Embed”.",
        "Ships dedicated Code Interpreter and Desktop (computer-use) SDKs."
      ],
      "pros": [
        "You want the most widely integrated agent sandbox, with a large ecosystem of examples",
        "You need Firecracker microVM isolation for untrusted multi-tenant code",
        "You want an open-source codebase you can self-host or run via BYOC",
        "You want ready-made code-interpreter and desktop SDKs"
      ],
      "cons": [
        "CPU-only — no GPUs",
        "Session length capped at 1 hour on the free plan and 24 hours on Pro",
        "SDKs documented for Python and JavaScript/TypeScript only",
        "Pro plan carries a $150/month platform fee"
      ],
      "verdict": "Open-source Firecracker microVM sandboxes with pause/resume, snapshots and forking; suited to code interpreters and short-to-medium-lived agent tasks, with a self-host path.",
      "peers": [
        "daytona",
        "modal",
        "vercel-sandbox",
        "fly-sprites",
        "cloudflare-sandbox",
        "runloop",
        "blaxel",
        "microsandbox"
      ],
      "sources": [
        "https://e2b.dev/",
        "https://e2b.dev/pricing",
        "https://docs.e2b.dev/",
        "https://docs.e2b.dev/sandbox.md",
        "https://docs.e2b.dev/faq/sandbox-lifetime.md",
        "https://github.com/e2b-dev/E2B",
        "https://github.com/e2b-dev/runtime",
        "https://e2b.dev/blog/series-a"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "daytona",
      "name": "Daytona",
      "company": "Daytona",
      "hq": {
        "city": "New York",
        "region": "New York",
        "country": "US",
        "source": "https://www.daytona.io/company/terms-of-service"
      },
      "kind": "hosted",
      "tier": 1,
      "alternatives": true,
      "website": "https://www.daytona.io",
      "docs": "https://www.daytona.io/docs",
      "pricing_url": "https://www.daytona.io/pricing",
      "github": "https://github.com/daytonaio/daytona",
      "tagline": "Stateful sandboxes for AI-generated code — Linux containers, Linux and Windows VMs, macOS and GPU sandboxes — with snapshots, volumes and preview URLs.",
      "isolation": "Isolated Linux containers (dedicated namespaces, resource limits) by default; VM sandboxes (Linux, Windows) run a full VM with their own kernel; GPU sandboxes are containers with exclusive GPU allocation",
      "isolation_short": "Containers or VMs",
      "isolation_class": "mixed",
      "open_source": false,
      "license": "Former public repo archived (AGPL-3.0)",
      "self_host": null,
      "byoc": true,
      "sdks": [
        "Python",
        "TypeScript",
        "Ruby",
        "Go",
        "Java"
      ],
      "cold_start": "Sub-90 ms sandbox creation (vendor claim)",
      "cold_start_short": "<90 ms",
      "max_session": "No hard maximum: auto-stop after 15 min idle by default (can be disabled), auto-archive after 7 days stopped, optional wall-clock TTL",
      "max_session_short": "Unlimited",
      "persistence": "Stateful sandboxes, snapshots, shared volumes, archive/recover and stop/start with state. Pause/resume, fork and hot snapshots on VM sandboxes",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Yes — up to 8 GPUs per sandbox (e.g. H100 from $2.27/hour, B200 $3.59/hour, preemptible list rates)",
      "networking": "Tiered egress: per-sandbox CIDR and domain allowlists or block-all, changeable at runtime; preview URLs for any port with token or signed URLs; SSH, VNC and web terminal",
      "pricing": "Per second: $0.0504 per vCPU-hour + $0.0162 per GiB-hour; storage $0.000108/GiB-hour after 5 GiB free; Windows $0.0858/vCPU-hour",
      "free_tier": "$200 free compute for new users; startup program up to $50k",
      "compliance": "States HIPAA, SOC 2 and GDPR compliance (Trust Center)",
      "regions": "US East, US West, EU Central, EU West, India; custom regions via BYOC",
      "best_for": "Long-lived, stateful agent workspaces that need many SDK languages, GPUs, or Windows and macOS sandboxes",
      "overview": [
        "Daytona began as a development-environment tool and now focuses on agent infrastructure and has become one of the most popular sandbox platforms. Sandboxes are **stateful by default** — they stop when idle and resume with files intact — which suits coding agents that return to the same workspace. It offers Linux containers, Linux and Windows VMs, macOS and GPU sandboxes from one API.",
        "Daytona’s SDK coverage (Python, TypeScript, Ruby, Go and Java) is the broadest among hosted providers, and BYOC lets you run sandboxes on your own runner nodes. In June 2026 Daytona announced that it was moving its production codebase to closed source; the public AGPL-3.0 repository is no longer maintained and was archived on October 3, 2026."
      ],
      "notable": [
        "Raised a $24M Series A in February 2026, led by FirstMark Capital.",
        "Moved core development to a private codebase in June 2026; the public repository was archived on October 3, 2026.",
        "Lists LangChain, SambaNova, Writer and n8n among customers."
      ],
      "pros": [
        "You need long-lived, stateful workspaces that survive between sessions",
        "You want SDKs beyond Python and TypeScript (Ruby, Go, Java)",
        "You need GPUs, Windows or macOS sandboxes from the same provider",
        "You want BYOC on your own compute"
      ],
      "cons": [
        "Production codebase closed source since June 2026 — the public repo is archived",
        "Default Linux sandboxes use container isolation; VM isolation requires VM sandboxes",
        "Fork and pause/resume limited to VM sandboxes"
      ],
      "verdict": "A stateful sandbox platform with a wide range of sandbox types, suited to coding-agent workspaces; its production codebase has been closed source since June 2026.",
      "peers": [
        "e2b",
        "fly-sprites",
        "modal",
        "runloop",
        "vercel-sandbox",
        "codesandbox",
        "northflank"
      ],
      "sources": [
        "https://www.daytona.io/",
        "https://www.daytona.io/pricing",
        "https://www.daytona.io/docs/en/sandboxes/",
        "https://www.daytona.io/docs/en/isolation/",
        "https://www.daytona.io/docs/en/network-limits/",
        "https://www.daytona.io/docs/en/regions/",
        "https://github.com/daytonaio/daytona",
        "https://www.daytona.io/dotfiles/updates/daytona-is-going-closed-source",
        "https://www.daytona.io/dotfiles/daytona-raises-24m-series-a-to-give-every-agent-a-computer",
        "https://www.daytona.io/dotfiles/daytona-raises-60m-series-b-to-build-the-agent-cloud"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "modal",
      "name": "Modal Sandboxes",
      "company": "Modal Labs",
      "hq": {
        "city": "New York",
        "region": "New York",
        "country": "US",
        "source": "https://modal.com/blog/general-availability-and-series-a-press-release"
      },
      "kind": "hosted",
      "tier": 1,
      "alternatives": true,
      "website": "https://modal.com/use-cases/sandboxes",
      "docs": "https://modal.com/docs/guide/sandbox",
      "pricing_url": "https://modal.com/pricing",
      "github": "https://github.com/modal-labs/modal-client",
      "tagline": "Sandboxes on Modal’s serverless compute platform for running untrusted code at very high scale, with GPUs, snapshots and tunnels.",
      "isolation": "gVisor or VMs with their own Linux kernel (Modal chooses if no runtime is set); GPU sandboxes use gVisor",
      "isolation_short": "gVisor or VM",
      "isolation_class": "gvisor",
      "open_source": false,
      "license": "Client SDK Apache-2.0",
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go"
      ],
      "cold_start": "Sub-second, including for large custom images (vendor claim)",
      "cold_start_short": "<1 s",
      "max_session": "5 minutes default, configurable up to 24 hours; snapshots carry state beyond that",
      "max_session_short": "Up to 24 h",
      "persistence": "Filesystem snapshots, directory snapshots, memory snapshots (alpha) and Modal Volumes",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Yes — Modal’s full GPU range (T4 to B300) at standard per-second rates, e.g. H100 $0.001097/s",
      "networking": "Outbound open by default; block all, CIDR allowlists or domain allowlists (beta), updatable at runtime; encrypted tunnels with public URLs and connect tokens",
      "pricing": "Per second at sandbox (non-preemptible) rates: $0.00003942 per physical core-second (1 core = 2 vCPU) + $0.00000667 per GiB-second; GPUs at standard rates. Team plan $250/month",
      "free_tier": "$30/month credits on Starter ($100/month on Team)",
      "compliance": "SOC 2 on all plans; HIPAA on Enterprise",
      "regions": "Region selection available at a price multiplier",
      "best_for": "High-concurrency and GPU-backed sandboxes — RL rollouts, evals and coding agents at scale",
      "overview": [
        "Modal is a serverless compute platform for Python-heavy AI workloads, and its Sandboxes product exposes the same infrastructure for running untrusted code. Its strengths are **scale** — Modal says its scheduler is tested with over a million concurrent sandboxes — and **GPUs**, which are available in sandboxes at the same rates as the rest of the platform.",
        "Sandboxes run on either gVisor or a VM runtime with its own Linux kernel. Snapshots (filesystem, directory and alpha memory snapshots) support fast restore and forking, which suits reinforcement learning and evaluation pipelines."
      ],
      "notable": [
        "Raised a $750M Series D led by Accel at a $15.75B valuation in October 2026.",
        "Says its scheduler is tested with over 1 million concurrent sandboxes.",
        "Lists Cognition, Linear and Ramp among sandbox customers."
      ],
      "pros": [
        "You need very high concurrency (RL, evals, batch agents)",
        "You need GPUs inside the sandbox",
        "You already use Modal for inference or training",
        "You want Go as well as Python and TypeScript SDKs"
      ],
      "cons": [
        "Sandbox compute billed at 3x Modal’s standard list rates",
        "Platform is not open source (client SDK is Apache-2.0); BYOC is not documented",
        "GPU sandboxes run only on the gVisor runtime, not the VM runtime"
      ],
      "verdict": "Sandboxes on Modal’s serverless platform with GPUs, snapshots and tunnels; suited to high-concurrency and GPU-backed workloads and to teams already on Modal.",
      "peers": [
        "e2b",
        "daytona",
        "runloop",
        "beam",
        "northflank",
        "morph"
      ],
      "sources": [
        "https://modal.com/use-cases/sandboxes",
        "https://modal.com/pricing",
        "https://modal.com/docs/guide/sandbox",
        "https://modal.com/docs/guide/sandbox-networking",
        "https://modal.com/docs/guide/sandbox-snapshots",
        "https://modal.com/blog/1-billion-annualized-run-rate-series-d"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "vercel-sandbox",
      "name": "Vercel Sandbox",
      "company": "Vercel",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://vercel.com/legal/modern-slavery-act"
      },
      "kind": "hosted",
      "tier": 1,
      "alternatives": true,
      "website": "https://vercel.com/sandbox",
      "docs": "https://vercel.com/docs/vercel-sandbox",
      "pricing_url": "https://vercel.com/docs/sandbox/pricing",
      "github": "https://github.com/vercel/sandbox",
      "tagline": "Firecracker microVM sandboxes on Vercel’s infrastructure, persistent by default and billed on active CPU, with fine-grained egress control.",
      "isolation": "Firecracker microVM with a dedicated kernel per sandbox",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": "SDK and CLI Apache-2.0",
      "self_host": null,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Python"
      ],
      "cold_start": "“Start in milliseconds” (vendor claim, no figure); snapshot resume faster than fresh start",
      "cold_start_short": "Milliseconds (claim)",
      "max_session": "5 min default; up to 45 min per session on Hobby and 24 hours on Pro/Enterprise; persistent sandboxes can be resumed indefinitely",
      "max_session_short": "24 h (Pro)",
      "persistence": "Persistent by default — filesystem auto-snapshotted on stop and restored on resume; manual snapshots; Drives (beta) and S3 mounts",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Firewall modes allow-all, deny-all or domain/CIDR rules, updatable at runtime; credential brokering injects headers so secrets never enter the sandbox; public URLs for up to 15 ports",
      "pricing": "Active CPU $0.128/hour (billed only while CPU is used) + provisioned memory $0.0212/GB-hour + $0.60 per million creations, plus storage and transfer (iad1 rates)",
      "free_tier": "Hobby: 5 active-CPU hours, 420 GB-hours memory and 5,000 creations per month",
      "compliance": "Runs on Vercel infrastructure with SOC 2 Type II (company-wide)",
      "regions": "19 regions, selectable per sandbox",
      "best_for": "Teams on Vercel and agent builders who want microVM isolation, active-CPU billing and credential brokering",
      "overview": [
        "Vercel Sandbox runs code in Firecracker microVMs on Vercel’s infrastructure and has been generally available since January 2026. Sandboxes are **persistent by default**: the filesystem is snapshotted on stop and restored on resume, so a coding agent can pick up where it left off.",
        "Two features stand out for agent security and cost: a **firewall with credential brokering** that injects secrets at the network layer so they never enter the sandbox, and **active-CPU billing** that charges CPU only while code is actually running — useful when agents spend most of their time waiting on a model."
      ],
      "notable": [
        "Generally available since January 30, 2026.",
        "Used in production by v0, Blackbox AI and RooCode per Vercel’s GA announcement."
      ],
      "pros": [
        "You deploy on Vercel or use the AI SDK",
        "You want secrets brokered outside the sandbox",
        "Your agents idle while waiting on models and you want active-CPU billing",
        "You want persistent-by-default microVMs"
      ],
      "cons": [
        "No GPU support documented",
        "Self-hosting and BYOC are not documented",
        "Hobby sessions capped at 45 minutes"
      ],
      "verdict": "A Firecracker microVM sandbox billed on active CPU, with fine-grained egress control and credential brokering; suited to teams already on Vercel.",
      "peers": [
        "e2b",
        "cloudflare-sandbox",
        "daytona",
        "fly-sprites",
        "freestyle",
        "codesandbox"
      ],
      "sources": [
        "https://vercel.com/docs/vercel-sandbox",
        "https://vercel.com/docs/sandbox/concepts",
        "https://vercel.com/docs/sandbox/pricing",
        "https://vercel.com/docs/sandbox/concepts/firewall",
        "https://vercel.com/changelog/vercel-sandboxes-ga",
        "https://github.com/vercel/sandbox"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "cloudflare-sandbox",
      "name": "Cloudflare Sandbox SDK",
      "company": "Cloudflare",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.cloudflare.com/about-overview/"
      },
      "kind": "hosted",
      "tier": 1,
      "alternatives": true,
      "website": "https://developers.cloudflare.com/sandbox/",
      "docs": "https://developers.cloudflare.com/sandbox/",
      "pricing_url": "https://developers.cloudflare.com/containers/pricing/",
      "github": "https://github.com/cloudflare/sandbox-sdk",
      "tagline": "A TypeScript SDK for Cloudflare Workers that runs untrusted code in Linux sandboxes backed by Cloudflare Containers and Durable Objects.",
      "isolation": "Each container instance is a microVM with its own kernel and network (per Cloudflare docs); Dynamic Workers (V8 isolates) available for JavaScript-only code",
      "isolation_short": "MicroVM (Containers)",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "SDK Apache-2.0 (runs only on Cloudflare)",
      "self_host": false,
      "byoc": false,
      "sdks": [
        "TypeScript (Workers)"
      ],
      "cold_start": "Container cold starts “often in the 1–3 second range” depending on image (vendor docs)",
      "cold_start_short": "1–3 s",
      "max_session": "No fixed cap; runs while its Durable Object is active, then up to a 6-hour inactivity timeout",
      "max_session_short": "No fixed cap",
      "persistence": "Ephemeral disk by default; container snapshots (up to 20 GB, 30 days), directory backups to R2, S3/R2 bucket mounts and Durable Object storage",
      "snapshots": true,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Egress controlled from Worker code: disable internet and intercept outbound HTTP to allow specific hosts and inject credentials; preview URLs routed through your Worker",
      "pricing": "Cloudflare Containers rates in 10 ms increments: $0.000020 per vCPU-second (active) + $0.0000025 per GiB-second + disk; requires the $5/month Workers Paid plan",
      "free_tier": "Workers Paid includes 25 GiB-hours memory, 375 vCPU-minutes and 200 GB-hours disk per month",
      "compliance": null,
      "regions": "Cloudflare’s global network",
      "best_for": "Teams building on Cloudflare Workers that want per-user sandboxes with egress and credentials controlled in code",
      "overview": [
        "Cloudflare’s Sandbox SDK lets a Worker create Linux sandboxes on demand, each a Durable Object paired with a Cloudflare Container. Containers and Sandboxes became generally available in April 2026, and Cloudflare documents each container instance as a microVM with its own kernel.",
        "Its distinctive design is that **your Worker sits in the network path**: you can run a sandbox with no internet and intercept only the outbound requests you allow, injecting credentials in Worker code. For JavaScript-only workloads, Dynamic Workers (V8 isolates) offer a much lighter option."
      ],
      "notable": [
        "Containers and Sandboxes generally available since April 13, 2026.",
        "Docs include guides for running Claude Code, Codex and other coding agents.",
        "SDK 1.0 is the current version; 0.x is marked legacy and receives bug and security fixes until December 31, 2026."
      ],
      "pros": [
        "You build on Cloudflare Workers and Durable Objects",
        "You want egress and credentials controlled in your own code",
        "You want global placement on Cloudflare’s network",
        "You want a lighter isolate option for JS-only code"
      ],
      "cons": [
        "SDK is used from Workers code (TypeScript); a Python SDK is not documented",
        "Container cold starts documented as often 1–3 seconds, depending on image",
        "Runs only on Cloudflare’s platform"
      ],
      "verdict": "A programmable sandbox SDK for Cloudflare Workers developers, driven from TypeScript Worker code on Cloudflare’s platform.",
      "peers": [
        "vercel-sandbox",
        "e2b",
        "fly-sprites",
        "freestyle",
        "daytona"
      ],
      "sources": [
        "https://developers.cloudflare.com/sandbox/",
        "https://developers.cloudflare.com/sandbox/concepts/lifetime/",
        "https://developers.cloudflare.com/sandbox/network/",
        "https://developers.cloudflare.com/containers/",
        "https://developers.cloudflare.com/containers/pricing/",
        "https://github.com/cloudflare/sandbox-sdk",
        "https://developers.cloudflare.com/containers/concepts/architecture/",
        "https://developers.cloudflare.com/sandbox/sdk/migrate/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "fly-sprites",
      "name": "Fly.io Sprites",
      "company": "Fly.io",
      "hq": {
        "city": "Chicago",
        "region": "Illinois",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/fly-io"
      },
      "kind": "hosted",
      "tier": 1,
      "alternatives": true,
      "website": "https://fly.io/sprites",
      "docs": "https://docs.fly.io/sprites",
      "pricing_url": "https://fly.io/sprites",
      "github": "https://github.com/superfly/sprites-go",
      "tagline": "Persistent, hardware-isolated Linux computers for AI agents with a durable 100 GB disk, checkpoints, and no compute charges while idle.",
      "isolation": "Dedicated microVM with hardware-level isolation (Fly docs); Fly.io says Sprites run on its Firecracker substrate",
      "isolation_short": "MicroVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": "SDKs MIT",
      "self_host": null,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Go",
        "Python",
        "Elixir"
      ],
      "cold_start": "Warm wake 100–500 ms with processes preserved; cold wake 1–2 s (vendor docs)",
      "cold_start_short": "0.1–2 s wake",
      "max_session": "No lifetime limit — a Sprite persists until destroyed and sleeps after ~30 s idle",
      "max_session_short": "Unlimited",
      "persistence": "100 GB durable disk per Sprite that survives sleep; copy-on-write checkpoints (~1 s) with restore; memory preserved on warm pause",
      "snapshots": true,
      "pause_resume": true,
      "gpu": false,
      "gpu_detail": "No GPU option documented",
      "networking": "HTTPS URL per Sprite (org-private by default); outbound open by default with optional DNS-based domain allow/deny policy; connectors broker API credentials",
      "pricing": "Per second, compute only while active: $0.0385 per CPU-hour + $0.021875 per GB-hour memory; storage billed while idle",
      "free_tier": "$1 trial credit for new Fly.io accounts, no card required; invoices under $5 are waived",
      "compliance": "Fly.io is SOC 2 Type 2; HIPAA BAA and DPA available (company-wide)",
      "regions": null,
      "best_for": "Long-running, stateful agent computers that should cost nothing while idle",
      "overview": [
        "Sprites, launched in January 2026, are Fly.io’s take on “computers for agents” rather than throwaway sandboxes. Each Sprite is a hardware-isolated microVM with a durable 100 GB disk, a public or private HTTPS URL and no lifetime limit. It sleeps after about 30 seconds idle and wakes on demand, and you pay only storage while it sleeps.",
        "Checkpoints take copy-on-write snapshots of the whole filesystem in about a second, so an agent can try something risky and roll back."
      ],
      "notable": [
        "Launched January 2026.",
        "Offers a remote MCP server and setups for Claude Managed Agents and Cursor Cloud Agents."
      ],
      "pros": [
        "You want a persistent computer per agent or per user",
        "Your agents are idle most of the time and you want to pay only for active compute",
        "You want fast checkpoints and rollback",
        "You want Elixir or Go SDKs"
      ],
      "cons": [
        "No GPUs",
        "Fixed 8 vCPU / 100 GB shape; memory platform-managed",
        "Self-hosting and BYOC are not documented"
      ],
      "verdict": "A persistent Linux computer per agent with a durable disk, checkpoints and no compute charges while idle; suited to long-running, stateful agents.",
      "peers": [
        "daytona",
        "e2b",
        "vercel-sandbox",
        "morph",
        "codesandbox"
      ],
      "sources": [
        "https://fly.io/sprites",
        "https://docs.fly.io/sprites",
        "https://docs.fly.io/sprites/concepts/lifecycle.md",
        "https://docs.fly.io/sprites/concepts/networking.md",
        "https://fly.io/compliance",
        "https://github.com/superfly/sprites-go",
        "https://fly.io/learn/sprites-vs-e2b-vs-modal-vs-daytona/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "northflank",
      "extra_vs": [
        "modal",
        "e2b"
      ],
      "name": "Northflank Sandboxes",
      "company": "Northflank",
      "hq": {
        "city": "London",
        "country": "GB",
        "source": "https://northflank.com/legal/terms"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": false,
      "website": "https://northflank.com/product/sandboxes",
      "docs": "https://northflank.com/docs/v1/application/sandboxes",
      "pricing_url": "https://northflank.com/pricing",
      "github": null,
      "tagline": "Sandboxes on Northflank’s Kubernetes-based platform — microVMs for CPU work, gVisor for GPUs — on Northflank’s cloud or in your own cloud account.",
      "isolation": "MicroVMs on Cloud Hypervisor with their own guest kernel for CPU workloads; gVisor for GPU workloads",
      "isolation_short": "Cloud Hypervisor microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": true,
      "sdks": [
        "JavaScript/TypeScript",
        "Python"
      ],
      "cold_start": "Under 1 s; sample “ready in 0.8 seconds” on a cached image (vendor claim)",
      "cold_start_short": "<1 s",
      "max_session": "No time limit; runs until deleted; idle sandboxes can be paused",
      "max_session_short": "Unlimited",
      "persistence": "Ephemeral by default; persistent and shared volumes, S3 mounts, managed databases and pause/resume",
      "snapshots": null,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Yes — L4 to B200 under gVisor (e.g. H100 $2.74/hour on Northflank Cloud) or on your own GPU nodes",
      "networking": "Private by default; network policies for ingress/egress by FQDN and IP range; mTLS service mesh; secrets injected at egress",
      "pricing": "Per second: $0.01667 per vCPU-hour + $0.00833 per GB-hour; BYOC $0.02083 per node vCPU-hour plus your cloud bill",
      "free_tier": "Free developer plan with up to $20/month of resources and one small BYOC cluster",
      "compliance": "SOC 2 Type II, HIPAA, ISO 27001",
      "regions": null,
      "best_for": "Teams that want sandboxes in their own cloud account (BYOC) alongside app hosting, databases and GPUs",
      "overview": [
        "Northflank is a full application platform that added sandboxes as a first-class workload. CPU sandboxes run in Cloud Hypervisor microVMs and GPU sandboxes under gVisor, either on Northflank’s cloud or inside your own AWS, Google Cloud, Azure, Oracle or bare-metal account via BYOC.",
        "On October 30, 2026 Northflank opened a security incident on its status page and advised customers to rotate all secrets stored on the platform. Its status page says it found no evidence that customer workloads or workload data were accessed, that forward-deployed and dedicated control planes were unaffected, and that a root cause analysis will follow."
      ],
      "notable": [
        "Northflank’s status page records a security incident opened on October 30, 2026; it advised customers to rotate all secrets and said users had reported unauthorised use of AI API keys supplied to the platform.",
        "Announced $22 million in seed and Series A funding in November 2024, led by Bain Capital Ventures and Vertex Ventures US."
      ],
      "pros": [
        "You need BYOC on any major cloud or bare metal",
        "You want GPUs and CPU sandboxes on one platform",
        "You also need app hosting, databases and pipelines",
        "You want low published per-vCPU rates"
      ],
      "cons": [
        "Security incident opened October 30, 2026; customers advised to rotate all secrets",
        "Not open source",
        "A general application platform; sandboxes are one of several workload types"
      ],
      "verdict": "An application platform offering microVM and gVisor sandboxes, suited to teams that want BYOC alongside app hosting, databases and GPUs; its status page records a security incident opened on October 30, 2026.",
      "peers": [
        "modal",
        "e2b",
        "daytona",
        "gke-agent-sandbox"
      ],
      "sources": [
        "https://northflank.com/product/sandboxes",
        "https://northflank.com/pricing",
        "https://northflank.com/docs/v1/application/sandboxes",
        "https://status.northflank.com/cmvv0rn5a092f14rutwccodpp",
        "https://northflank.com/blog/northflank-raises-22m-to-make-kubernetes-work-for-your-developers-ship-workloads-not-infrastructure",
        "https://northflank.com/security"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "blaxel",
      "extra_vs": [
        "e2b",
        "fly-sprites"
      ],
      "name": "Blaxel",
      "company": "Blaxel (acquired by Baseten)",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/blaxel"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": false,
      "website": "https://blaxel.ai",
      "docs": "https://docs.blaxel.ai",
      "pricing_url": "https://blaxel.ai/pricing",
      "github": "https://github.com/blaxel-ai",
      "tagline": "Perpetual microVM sandboxes for AI agents that drop to standby when idle and resume with memory intact, plus volumes, batch jobs and MCP hosting.",
      "isolation": "Hardware-isolated microVM per sandbox; root filesystem held in memory; hypervisor not named",
      "isolation_short": "MicroVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": "SDKs MIT / Apache-2.0",
      "self_host": null,
      "byoc": true,
      "sdks": [
        "TypeScript",
        "Python",
        "Go"
      ],
      "cold_start": "Resume from standby in under 25 ms (vendor claim); new sandboxes in under a second",
      "cold_start_short": "<25 ms resume",
      "max_session": "Designed to persist indefinitely in standby; workloads from seconds to multiple days",
      "max_session_short": "Unlimited (standby)",
      "persistence": "Automatic standby after ~5–15 s idle snapshots memory, processes and filesystem; forkable snapshots; volumes and a shared Agent Drive",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Preview URLs with stable names; outbound via proxy, firewall rules or dedicated egress IPs; custom domains",
      "pricing": "Per second on allocated memory: $0.0000115 per GB-second (~$0.041/GB-hour) while active; standby billed as snapshot storage ($0.20/GB-month)",
      "free_tier": "Up to $200 in credits, no card required",
      "compliance": "SOC 2, ISO 27001, GDPR, CCPA; HIPAA with BAA as a paid add-on",
      "regions": "US and Europe",
      "best_for": "Always-available agents that sit idle most of the time and need near-instant resume",
      "overview": [
        "Blaxel’s sandboxes are built around **standby**: after a few seconds of inactivity a sandbox snapshots its memory, processes and filesystem, stops consuming compute, and resumes in tens of milliseconds when called again. That makes “one sandbox per user, forever” affordable.",
        "Blaxel also offers batch jobs, MCP server hosting and agent hosting. Baseten announced its acquisition of Blaxel on September 10, 2026; Blaxel says the platform continues to operate."
      ],
      "notable": [
        "Acquisition by Baseten announced September 10, 2026.",
        "Raised a $7.3M seed led by First Round Capital (YC X25).",
        "Native sandbox support in the OpenAI Agents SDK announced April 2026."
      ],
      "pros": [
        "You want a persistent sandbox per user without paying for idle compute",
        "You need very fast resume from standby",
        "You want Go as well as Python and TypeScript SDKs"
      ],
      "cons": [
        "Acquired by Baseten in September 2026",
        "GPU support not documented",
        "Platform is not open source (SDKs are MIT / Apache-2.0)"
      ],
      "verdict": "A microVM sandbox platform built around automatic standby, suited to always-on, mostly idle agents; Blaxel was acquired by Baseten in September 2026.",
      "peers": [
        "fly-sprites",
        "e2b",
        "daytona",
        "freestyle",
        "vercel-sandbox"
      ],
      "sources": [
        "https://blaxel.ai",
        "https://blaxel.ai/pricing",
        "https://blaxel.ai/blog",
        "https://docs.blaxel.ai/Sandboxes/Overview",
        "https://github.com/blaxel-ai",
        "https://blaxel.ai/blog/blaxel-is-joining-baseten-to-build-the-future-of-agentic-infrastructure",
        "https://www.baseten.co/blog/blaxel-is-joining-baseten-to-build-the-future-of-agentic-cloud/",
        "https://blaxel.ai/blog/Blaxel-Raises-7-3M-Seed-Round-led-by-First-Round-to-Build-Cloud-Infrastructure-for-the-AI-Agent-Eco-23247e47b1ea8067b923d998364e3ced",
        "https://blaxel.ai/blog/native-support-for-blaxel-sandboxes-in-the-new-open-ai-agents-sdk"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "codesandbox",
      "extra_vs": [
        "e2b"
      ],
      "name": "Together Code Sandbox",
      "company": "Together AI (acquired CodeSandbox)",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.together.ai/terms-of-service"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": false,
      "website": "https://www.together.ai/code-sandbox",
      "docs": "https://docs.together.ai/docs/together-code-sandbox",
      "pricing_url": "https://www.together.ai/pricing",
      "github": "https://github.com/togethercomputer/together-sandbox",
      "tagline": "Together AI’s microVM sandbox service, built on the CodeSandbox acquisition, with snapshot-based starts, hibernation, cloning and preview URLs.",
      "isolation": "MicroVMs (Together and CodeSandbox documentation)",
      "isolation_short": "MicroVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": "Client SDK MIT",
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "TypeScript"
      ],
      "cold_start": "Start from memory snapshot in 500 ms (P95); clone in under 1 s; 2.7 s (P95) cold start (vendor claims)",
      "cold_start_short": "0.5 s (snapshot)",
      "max_session": "Runs until terminated unless a TTL is set",
      "max_session_short": "Until terminated",
      "persistence": "Hibernation snapshots disk and memory and resumes exact state; hibernated sandboxes can be forked; snapshots from Dockerfiles",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Preview URLs for exposed ports; egress controls not documented",
      "pricing": "$0.0446 per vCPU-hour + $0.0149 per GiB-hour; custom pricing for large deployments",
      "free_tier": null,
      "compliance": null,
      "regions": null,
      "best_for": "AI app builders and coding products that need hibernatable, forkable dev-environment VMs with preview URLs",
      "overview": [
        "Together AI acquired CodeSandbox in December 2024 and relaunched its microVM infrastructure as Together Code Sandbox in May 2025. It inherits years of experience running browser-based development environments: memory-snapshot starts, hibernation that restores exact state, and fast cloning.",
        "New integrations use the `together-sandbox` SDK; the original `@codesandbox/sdk` remains documented for existing users. Access is enabled through Together’s sales team."
      ],
      "notable": [
        "CodeSandbox acquired by Together AI, announced December 2024.",
        "Relaunched as Together Code Sandbox in May 2025 alongside Together Code Interpreter."
      ],
      "pros": [
        "You build an AI coding product that needs dev-environment-style VMs",
        "You want hibernate-and-resume with memory preserved",
        "You already use Together AI for inference"
      ],
      "cons": [
        "Access must be enabled per organisation via Together’s sales team",
        "Egress controls and compliance not documented",
        "Two SDKs: together-sandbox for new integrations, @codesandbox/sdk for existing users"
      ],
      "verdict": "MicroVM development-environment sandboxes built on CodeSandbox’s infrastructure, suited to teams already working with Together AI.",
      "peers": [
        "daytona",
        "e2b",
        "freestyle",
        "vercel-sandbox",
        "fly-sprites"
      ],
      "sources": [
        "https://www.together.ai/code-sandbox",
        "https://docs.together.ai/docs/together-code-sandbox",
        "https://www.together.ai/pricing",
        "https://www.together.ai/blog/code-sandbox-code-interpreter",
        "https://github.com/togethercomputer/together-sandbox",
        "https://www.together.ai/blog/codesandbox-acquisition-together-code-interpreter",
        "https://codesandbox.io/blog/joining-together-ai-introducing-codesandbox-sdk"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "morph",
      "extra_vs": [
        "e2b"
      ],
      "name": "Morph Cloud",
      "company": "Morph Labs",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": false,
      "website": "https://cloud.morph.so",
      "docs": "https://cloud.morph.so/docs/documentation/overview",
      "pricing_url": "https://cloud.morph.so/web/pricing",
      "github": "https://github.com/morph-labs",
      "tagline": "VM-based cloud computers for AI agents built on “Infinibranch” snapshot, branch and restore, plus persistent Devboxes.",
      "isolation": "Full VMs; hypervisor not named",
      "isolation_short": "VM",
      "isolation_class": "vm",
      "open_source": false,
      "license": "SDKs Apache-2.0",
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "TypeScript"
      ],
      "cold_start": "Snapshot, branch and restore entire environments in under 250 ms (vendor claim)",
      "cold_start_short": "<250 ms (restore)",
      "max_session": "Configurable TTL per instance; Devboxes can disable auto-sleep to run like a persistent VPS",
      "max_session_short": "Configurable",
      "persistence": "Snapshots, branching many instances from one state, pause/resume preserving memory and processes, volumes, wake-on-request",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Public HTTP URLs with optional API-key auth; SSH and reverse tunnels; egress controls not documented",
      "pricing": "Morph Compute Units at $0.05 each (1 MCU = 1 vCPU-hour + 4 GB RAM-hours + 16 GB disk-hours); Team $40/month incl. 1,000 MCUs",
      "free_tier": "Developer plan $0/month with no included compute units; MCUs are purchased separately",
      "compliance": null,
      "regions": null,
      "best_for": "Workloads that branch a running environment many times — parallel search, RL and evaluation rollouts",
      "overview": [
        "Morph Cloud’s core idea is **branching**: snapshot a running VM — memory, processes and disk — and start many copies from that exact state in a fraction of a second. That suits tree search, best-of-N agent attempts and RL rollouts, where you want to explore many paths from one prepared environment.",
        "Devboxes extend the same infrastructure to persistent development workspaces."
      ],
      "notable": [
        "Also sells GitHub Actions runners billed in compute units."
      ],
      "pros": [
        "You run many parallel attempts from one environment state",
        "You want pause/resume that keeps running processes",
        "You like a single blended compute unit for billing"
      ],
      "cons": [
        "Hypervisor and egress controls not documented",
        "No GPU support documented",
        "Compliance attestations not documented"
      ],
      "verdict": "Purpose-built for branching workloads; a specialist pick for search, RL and evals.",
      "peers": [
        "modal",
        "e2b",
        "freestyle",
        "daytona"
      ],
      "sources": [
        "https://cloud.morph.so",
        "https://cloud.morph.so/web/pricing",
        "https://cloud.morph.so/docs/documentation/overview",
        "https://cloud.morph.so/docs/documentation/instances/basic-lifecycle",
        "https://cloud.morph.so/docs/documentation/instances/pause-resume"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "beam",
      "extra_vs": [
        "modal"
      ],
      "name": "Beam Sandboxes",
      "company": "Beam",
      "hq": {
        "city": "New York",
        "region": "New York",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/beam"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": false,
      "website": "https://www.beam.cloud",
      "docs": "https://docs.beam.cloud/v2/sandbox/overview",
      "pricing_url": "https://www.beam.cloud/pricing",
      "github": "https://github.com/beam-cloud/beta9",
      "tagline": "Serverless AI compute platform whose sandboxes are gVisor-isolated containers with memory snapshots, optional GPUs and an open-source, self-hostable runtime.",
      "isolation": "gVisor + runc containers (Beam’s own guide recommends microVMs for adversarial multi-tenant code)",
      "isolation_short": "gVisor",
      "isolation_class": "gvisor",
      "open_source": true,
      "license": "AGPL-3.0 (beta9 runtime)",
      "self_host": true,
      "byoc": true,
      "sdks": [
        "Python",
        "TypeScript"
      ],
      "cold_start": "About half a second; 100 concurrent sandboxes in 0.55 s (vendor claim)",
      "cold_start_short": "~0.5 s",
      "max_session": "600 s default lifetime, extendable or disabled",
      "max_session_short": "Configurable",
      "persistence": "Filesystem snapshots, memory snapshots, pause/restore, volumes and durable disks",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Yes — GPUs attachable to sandboxes (e.g. RTX 4090 $0.69/hour)",
      "networking": "Public preview URLs via expose_port (add your own auth); runtime egress policies",
      "pricing": "Billed by the millisecond: performance tier $0.05/vCPU-hour + $0.01/GiB-hour; cheaper shared and spot tiers; plans from $20/month",
      "free_tier": "Developer plan $20/month includes $20 of credit",
      "compliance": "SOC 2 Type II",
      "regions": null,
      "best_for": "Teams that want an open-source, self-hostable sandbox runtime with GPU access",
      "overview": [
        "Beam is a serverless GPU platform whose sandboxes run as gVisor-isolated containers with millisecond billing, memory snapshots and attachable GPUs. Its runtime, **beta9**, is open source under AGPL-3.0 and can be self-hosted on Kubernetes, with a BYOC option for running in your own cloud.",
        "In its own guide, Beam describes gVisor as lighter than a full microVM and recommends microVMs for adversarial multi-tenant code from untrusted users."
      ],
      "notable": [
        "Backed by Y Combinator and others listed on its site."
      ],
      "pros": [
        "You want GPUs inside sandboxes at low rates",
        "You want an open-source runtime you can self-host",
        "You want spot and shared tiers for cheap batch work"
      ],
      "cons": [
        "gVisor rather than microVM isolation",
        "Runtime is licensed under AGPL-3.0",
        "Preview URLs need your own authentication"
      ],
      "verdict": "GPU-capable gVisor sandboxes with an open-source, self-hostable runtime, suited to teams that want low-cost GPU access; Beam’s own guide recommends microVMs for adversarial multi-tenant code.",
      "peers": [
        "modal",
        "northflank",
        "daytona",
        "koyeb"
      ],
      "sources": [
        "https://www.beam.cloud",
        "https://www.beam.cloud/pricing",
        "https://docs.beam.cloud/v2/sandbox/overview",
        "https://docs.beam.cloud/v2/sandbox/snapshots",
        "https://github.com/beam-cloud/beta9",
        "https://www.beam.cloud/blog/how-to-self-host-code-sandbox"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "freestyle",
      "extra_vs": [
        "daytona",
        "e2b"
      ],
      "name": "Freestyle",
      "company": "Freestyle",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/freestyle"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": false,
      "website": "https://www.freestyle.sh",
      "docs": "https://www.freestyle.sh/docs",
      "pricing_url": "https://www.freestyle.sh/pricing",
      "github": "https://github.com/freestyle-sh",
      "tagline": "Full Linux VMs for AI agents with live forking of running VMs, memory-preserving hibernation, VPCs and nested virtualization.",
      "isolation": "Full Linux VMs with hardware virtualization on Freestyle’s own bare metal; supports nested virtualization, FUSE and eBPF",
      "isolation_short": "VM",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "TypeScript"
      ],
      "cold_start": "Ready in 65 ms from API request; p99 under 400 ms; resume under 100 ms (vendor claims)",
      "cold_start_short": "65 ms",
      "max_session": "Can run indefinitely with idle timeout disabled (paid plans)",
      "max_session_short": "Unlimited (paid)",
      "persistence": "Live forking of running VMs including memory; hibernate with memory preserved; snapshots",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Default-deny network access with per-VM rules; VPCs, firewall, WireGuard/Tailscale tunnels; custom-domain previews with HTTPS",
      "pricing": "$0.04032 per vCPU-hour + $0.0129 per GiB-hour; Hobby $50/month and Pro $500/month credited against usage",
      "free_tier": "Free plan: 200 vCPU-hours and 400 GiB-hours per month (no persistent VMs)",
      "compliance": null,
      "regions": null,
      "best_for": "Long-running coding agents that need full VMs — Docker, nested virtualization — with fork and hibernate",
      "overview": [
        "Freestyle offers full Linux VMs rather than containers or minimal microVMs, running on its own bare-metal hardware. That means agents can run Docker, nested VMs, FUSE and eBPF — things many sandboxes block. Running VMs can be live-forked with memory intact and hibernated so only storage is billed.",
        "Default-deny networking with per-VM rules and VPC support make it a security-conscious choice for long-running agents. Freestyle is one of the migration targets named in Runloop’s 2026 wind-down guides."
      ],
      "notable": [
        "Named as a migration target in Runloop’s wind-down guides, with credits for Runloop customers.",
        "Backed by Floodgate, Y Combinator and Two Sigma Ventures, per its site."
      ],
      "pros": [
        "Your agent needs Docker, nested virtualization or kernel features",
        "You want to fork a running VM including memory",
        "You want default-deny networking and VPCs"
      ],
      "cons": [
        "Only a TypeScript SDK is documented",
        "Compliance attestations not documented",
        "GPU support not documented"
      ],
      "verdict": "Full Linux VMs with live forking, memory-preserving hibernation and nested virtualization; suited to long-running coding agents that need Docker or a full VM.",
      "peers": [
        "daytona",
        "e2b",
        "fly-sprites",
        "morph",
        "codesandbox"
      ],
      "sources": [
        "https://www.freestyle.sh",
        "https://www.freestyle.sh/pricing",
        "https://www.freestyle.sh/docs",
        "https://www.freestyle.sh/products/vms",
        "https://github.com/freestyle-sh",
        "https://docs.runloop.ai/docs/transitions/overview"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "koyeb",
      "name": "Koyeb Sandboxes",
      "company": "Koyeb (joining Mistral AI)",
      "hq": {
        "city": "Boulogne-Billancourt",
        "region": "Île-de-France",
        "country": "FR",
        "source": "https://www.koyeb.com/docs/legal/terms"
      },
      "kind": "hosted",
      "tier": 3,
      "alternatives": false,
      "website": "https://www.koyeb.com",
      "docs": "https://www.koyeb.com/docs/sandboxes",
      "pricing_url": "https://www.koyeb.com/pricing",
      "github": "https://github.com/koyeb/koyeb-sandbox-sdk-js",
      "tagline": "Ephemeral microVM sandboxes on Koyeb’s serverless platform for AI-generated code, with CPU or GPU and scale-to-zero.",
      "isolation": "MicroVMs on bare-metal servers",
      "isolation_short": "MicroVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": "SDKs Apache-2.0",
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript"
      ],
      "cold_start": "No sandbox-specific figure; platform cold starts “as low as 250 ms” (vendor claim)",
      "cold_start_short": null,
      "max_session": "Auto-deletion configurable from 60 seconds to 24 hours",
      "max_session_short": "Up to 24 h",
      "persistence": "Ephemeral by design with scale-to-zero deep sleep",
      "snapshots": null,
      "pause_resume": null,
      "gpu": true,
      "gpu_detail": "Yes — CPU and GPU sandboxes (platform GPUs from $0.50/hour to B200 at $5.50/hour)",
      "networking": "expose_port returns a public URL; egress controls not documented",
      "pricing": "Usage-based on paid plans (Pro $29/month); sandboxes from $0.000001 per second (launch blog)",
      "free_tier": "No sandbox free tier stated",
      "compliance": "SOC 2 and ISO 27001 (platform-wide)",
      "regions": "10 regions across 3 continents (platform)",
      "best_for": "Teams already on Koyeb or in the Mistral ecosystem that want GPU-capable ephemeral sandboxes",
      "overview": [
        "Koyeb added microVM sandboxes to its serverless platform in public preview in November 2025, with CPU and GPU options and scale-to-zero. On February 17, 2026 Koyeb announced a definitive agreement to join Mistral AI, saying its platform, technology and team would be brought to Mistral Compute."
      ],
      "notable": [
        "Koyeb announced a definitive agreement to join Mistral AI on February 17, 2026.",
        "Sandboxes launched in public preview in November 2025."
      ],
      "pros": [
        "You want GPU-capable microVM sandboxes",
        "You already deploy on Koyeb",
        "You are aligned with the Mistral ecosystem"
      ],
      "cons": [
        "In public preview",
        "Persistence and egress controls not documented",
        "Has agreed to join Mistral AI (announced February 2026)"
      ],
      "verdict": "GPU-capable microVM sandboxes in public preview on Koyeb’s serverless platform, suited to teams already deploying on Koyeb; Koyeb has agreed to join Mistral AI.",
      "peers": [
        "modal",
        "beam",
        "e2b",
        "northflank"
      ],
      "sources": [
        "https://www.koyeb.com/pricing",
        "https://www.koyeb.com/blog/koyeb-sandboxes-fast-scalable-fully-isolated-environments-for-ai-agents",
        "https://www.koyeb.com/docs/sandboxes",
        "https://www.koyeb.com/docs/sandboxes/sandbox-lifecycle",
        "https://www.koyeb.com/blog/koyeb-is-joining-mistral-ai-to-build-the-future-of-ai-infrastructure"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "runloop",
      "name": "Runloop",
      "company": "Runloop AI",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://runloop.ai/media/runloop-raises-7m-seed-round-to-bring-enterprise-grade-infrastructure-to-ai-coding-agents"
      },
      "kind": "hosted",
      "tier": 3,
      "alternatives": true,
      "status": "winding-down",
      "status_note": "Runloop’s documentation states that Runloop is winding down its services, with transition support for existing customers running through **November 13, 2026 at 11:59 PM PT**, a date that has now passed. Its docs include migration guides to Tensorlake, Modal and Freestyle. We do not recommend it for new projects.",
      "website": "https://runloop.ai",
      "docs": "https://docs.runloop.ai",
      "pricing_url": "https://runloop.ai/pricing",
      "github": "https://github.com/runloopai",
      "tagline": "Managed microVM “Devboxes” for AI coding agents with blueprints, snapshots, suspend/resume and benchmarking tools — now winding down.",
      "isolation": "MicroVM plus container on a custom bare-metal hypervisor",
      "isolation_short": "MicroVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": true,
      "sdks": [
        "Python",
        "TypeScript"
      ],
      "cold_start": "10 GB image in under 2 s (vendor claim)",
      "max_session": "1 hour default, configurable",
      "persistence": "Disk snapshots, suspend/resume (disk only), Dockerfile-based blueprints",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Optional GPU listed in docs; no types or prices published",
      "networking": "Per-devbox egress policies, HTTPS tunnels, agent gateway for credentials",
      "pricing": "$0.108 per CPU-hour + $0.0252 per GB-hour (service winding down)",
      "free_tier": null,
      "compliance": "SOC 2 Type II",
      "regions": null,
      "best_for": "Existing customers planning a migration — not suitable for new projects",
      "overview": [
        "Runloop offered microVM Devboxes for coding agents, with blueprints, snapshots and agent-benchmarking tools. Its documentation says Runloop is winding down its services, with transition support for existing customers through November 13, 2026.",
        "If you are migrating, Runloop’s own guides point to **Tensorlake**, **[Modal](/sandboxes/modal/)** and **[Freestyle](/sandboxes/freestyle/)**. For devbox-style persistent workspaces, also consider [Daytona](/sandboxes/daytona/), [E2B](/sandboxes/e2b/) and [Fly.io Sprites](/sandboxes/fly-sprites/). See [Runloop alternatives](/alternatives/runloop/)."
      ],
      "notable": [
        "Documentation states that the service is winding down; transition support for existing customers ran through November 13, 2026.",
        "Runloop’s Freestyle migration guide lists an offer for Runloop customers of $1,000 in Freestyle credits plus one month of the Hobby plan."
      ],
      "pros": [
        "Purpose-built coding-agent devboxes with evaluation tooling (historical)"
      ],
      "cons": [
        "Service is winding down; transition support ran through November 13, 2026",
        "Not suitable for new projects"
      ],
      "verdict": "Winding down: Runloop’s documentation says transition support for existing customers ran through November 13, 2026.",
      "peers": [
        "freestyle",
        "modal",
        "daytona",
        "e2b",
        "fly-sprites",
        "morph"
      ],
      "alt_intro": "Runloop is winding down its services, so each entry notes how closely it matches Runloop’s devbox model and what a migration involves.",
      "sources": [
        "https://runloop.ai",
        "https://runloop.ai/pricing",
        "https://docs.runloop.ai/docs/transitions/overview",
        "https://docs.runloop.ai/docs/devboxes/start-stop",
        "https://docs.runloop.ai/docs/transitions/runloop-to-freestyle"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "riza",
      "name": "Riza",
      "company": "Riza",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://riza.io/company"
      },
      "kind": "hosted",
      "tier": 3,
      "alternatives": true,
      "status": "discontinued",
      "status_label": "hosted API discontinued",
      "status_note": "Riza announced on July 28, 2025 that it would shut down its hosted Code Interpreter API on **October 1, 2025**, and that its self-hosted container would keep working for existing licence holders. On August 11, 2025 Riza pointed its users to Daytona. We found no later announcement on Riza’s blog.",
      "website": "https://riza.io",
      "docs": null,
      "pricing_url": null,
      "github": null,
      "tagline": "A hosted code-interpreter API for running LLM-generated code; Riza announced that the hosted service would shut down on October 1, 2025.",
      "isolation": null,
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [],
      "cold_start": null,
      "max_session": null,
      "persistence": null,
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": null,
      "pricing": null,
      "free_tier": null,
      "compliance": null,
      "regions": null,
      "best_for": "Historical reference — Riza announced the end of its hosted service",
      "overview": [
        "Riza provided a simple API for running untrusted, LLM-generated code. It announced on July 28, 2025 that the hosted API and dashboard would shut down on October 1, 2025, and partnered with Daytona to offer migration credits.",
        "For a hosted code interpreter today, see [sandboxes for code interpreters](/use-cases/code-interpreter/) and [Riza alternatives](/alternatives/riza/)."
      ],
      "notable": [
        "Shutdown of the hosted API on October 1, 2025 announced by Riza on July 28, 2025."
      ],
      "pros": [],
      "cons": [
        "Riza announced the shutdown of its hosted service for October 1, 2025"
      ],
      "verdict": "Riza announced that its hosted Code Interpreter API would shut down on October 1, 2025; for new work, see the active code-interpreter sandboxes.",
      "peers": [
        "e2b",
        "daytona",
        "azure-dynamic-sessions",
        "aws-agentcore",
        "openai-code-interpreter"
      ],
      "alt_intro": "Riza announced the shutdown of its hosted API, so these are the closest active options for running LLM-generated code.",
      "sources": [
        "https://riza.io/blog/shutting-down-hosted-code-interpreter",
        "https://riza.io/blog/partnering-with-daytona",
        "https://riza.io"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "boat",
      "name": "boat",
      "company": "Boat",
      "hq": {
        "city": "Paris",
        "country": "FR",
        "source": "https://www.ycombinator.com/companies/boat"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://boat.dev",
      "docs": "https://docs.boat.dev",
      "pricing_url": "https://docs.boat.dev/pricing",
      "github": null,
      "tagline": "Persistent Ubuntu VMs for AI agents in EU data centres, with SSH, Docker, disk snapshots, forking, a 60 fps virtual desktop and per-second billing from a $20/month plan.",
      "isolation": "Full virtual machine per sandbox with its own kernel, disk and IP address; hypervisor not named",
      "isolation_short": "Full VM, own kernel",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "CLI",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "No fixed cap on paid plans: auto-stop defaults to 1 hour after start or resume, can be set up to 30 days or turned off; free trial is limited to 2 hours of auto-stop",
      "max_session_short": "No cap (auto-stop 1 h default)",
      "persistence": "Disk is snapshotted every minute while running and on every stop; stopping pauses billing and keeps the disk for at least 14 days; `boat fork` clones a sandbox from a snapshot; running processes do not come back after a resume",
      "snapshots": true,
      "pause_resume": true,
      "gpu": false,
      "gpu_detail": "No - \"No size has a GPU\" (pricing docs); the homepage says \"Graphics run on the CPU\"",
      "networking": "Outbound open by default (public IPv6 plus outbound IPv4, some sandboxes IPv6-only); inbound refused except SSH and boat ports, ports exposed via `host <port>` public URLs or ufw; no API-level network rules, and the default user has passwordless sudo so in-sandbox firewall rules can be removed by the agent; egress is not billed",
      "pricing": "Per second while running: small 2 vCPU/4 GB $0.018/h, default 4 vCPU/8 GB $0.036/h, large 8 vCPU/16 GB $0.072/h, xlarge 16 vCPU/32 GB $0.200/h ($100 plan or higher). Monthly plans of $20 (100 concurrent sandboxes, 555 h of default VM time), $100 (300, 2,777 h), $500 (1,000, 13,888 h) and $2,000 (2,000, 55,555 h); included time does not roll over. Saved snapshots beyond the plan allowance $0.056/day each; extra starts $0.03-$0.07 each",
      "free_tier": "7-day free trial (card required) with 25 compute hours, 2 concurrent sandboxes and auto-stop capped at 2 hours; no $0 plan",
      "compliance": "SOC 2 listed as \"in progress\", not yet certified; zero-data-retention option; measured uptime 99.95%",
      "regions": "Germany, Finland and France only (no US or Asia regions); snapshot copies may be stored outside the EU",
      "best_for": "Long-lived EU-hosted VMs that an agent keeps for days, with SSH, Docker and a desktop",
      "overview": [
        "boat sells full Ubuntu virtual machines (\"sandboxes\") for AI agents, from a Y Combinator-backed company; the product was named Box until 15 September 2026 (box.ascii.dev and docs.ascii.dev now redirect to boat.dev, and the SDK docs carry a migration guide from the legacy box SDKs). Each sandbox is its own VM with its own kernel, disk and IP; the hypervisor is not named. You reach it over SSH/SCP, the `boat` CLI, an HTTP API or Python and TypeScript SDKs, and Docker, VS Code, Chrome, Node, Rust, Bun and coding agents such as Claude Code, Codex and OpenCode are preinstalled. `boat desktop` streams a 60 fps virtual desktop rendered on the CPU.",
        "Disk is snapshotted every minute and on stop; a stopped sandbox keeps its disk for at least 14 days and can be resumed or forked, but memory and processes are not preserved. Billing is per second with a $20/month minimum plan that includes 555 hours of the default 4 vCPU/8 GB size, and the company publishes its own price comparison against other providers (a vendor claim). Machines are in Germany, Finland and France only; SOC 2 is in progress rather than certified. Starts per minute, hour and day are rate-limited by plan, which matters for burst-heavy workloads."
      ],
      "notable": [
        "Backed by Y Combinator (Fall 2026 batch); the YC profile lists the company as founded in 2024",
        "Renamed from Box to boat on 15 September 2026: box.ascii.dev and docs.ascii.dev redirect to boat.dev, and the status page remains at status.ascii.dev",
        "List price of $0.036/hour for a 4 vCPU/8 GB VM, which boat claims is the lowest per-VM-hour price among the providers it checked"
      ],
      "pros": [
        "You want a persistent full VM with SSH, Docker and root that an agent can keep for days or weeks",
        "You need EU-only compute (Germany, Finland, France) for sandbox workloads",
        "You want predictable per-second VM pricing with a low monthly minimum and no egress charges",
        "You want a streamed virtual desktop without paying for a GPU"
      ],
      "cons": [
        "No US or Asian regions, and snapshot copies may be stored outside the EU",
        "SOC 2 is \"in progress\", not certified",
        "No GPUs on any size",
        "Network policy cannot be set through the API, and the default user has passwordless sudo, so in-VM firewall rules can be removed by the agent",
        "Sandbox starts are rate-limited per minute, hour and day by plan, with no queueing"
      ],
      "verdict": "A low-priced EU-hosted full VM for agents that need persistence and a desktop, with no US or Asia region, plan-based start rate limits and SOC 2 still in progress.",
      "peers": [
        "exe-dev",
        "upstash-box",
        "fly-sprites",
        "tensorlake",
        "e2b"
      ],
      "sources": [
        "https://boat.dev",
        "https://docs.boat.dev",
        "https://docs.boat.dev/pricing",
        "https://docs.boat.dev/security",
        "https://docs.boat.dev/compliance",
        "https://docs.boat.dev/limits.md",
        "https://docs.boat.dev/long-running-tasks.md",
        "https://docs.boat.dev/sdks/overview",
        "https://box.ascii.dev",
        "https://docs.ascii.dev/llms.txt",
        "https://www.ycombinator.com/companies/boat",
        "https://docs.boat.dev/migrating-from-box",
        "https://docs.boat.dev/free-trial"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "deno-sandbox",
      "name": "Deno Sandbox",
      "company": "Deno Land (team joining Cloudflare)",
      "hq": {
        "city": "San Diego",
        "region": "California",
        "country": "US",
        "source": "https://docs.deno.com/deploy/privacy_policy/"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://deno.com/deploy/sandbox",
      "docs": "https://docs.deno.com/sandbox/",
      "pricing_url": "https://deno.com/deploy/sandbox#sandbox-pricing",
      "github": "https://github.com/denoland/sandbox-py",
      "tagline": "Linux microVMs on Deno Deploy with host allow-lists and proxy-injected secrets; launched in beta in February 2026 and, as of November 2026, no longer available to new customers.",
      "isolation": "Linux microVM per sandbox, \"isolated at the hypervisor level\"; product page says Firecracker, the security docs do not name the hypervisor",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Python",
        "CLI",
        "REST API"
      ],
      "cold_start": "Product page FAQ: \"Under 200ms\"; launch post: boot time under 1 second (vendor claims)",
      "cold_start_short": "<200 ms (claim)",
      "max_session": "Session-bound by default; a duration timeout keeps it alive after disconnect and can be extended, with the About page and launch post giving a 30-minute lifetime cap (extendable on demand); longer-lived work is meant to be promoted to a Deno Deploy app",
      "max_session_short": "30 min (extendable)",
      "persistence": "VMs are ephemeral and boot from a clean image; read-write volumes (300 MB-20 GB, one sandbox at a time, Chicago region only) survive sandbox teardown, and read-only snapshots of a volume can be used as the root filesystem for many sandboxes at once; pause/resume of a running VM is not documented",
      "snapshots": true,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not mentioned in the docs or on the product page",
      "networking": "Outbound goes through a proxy enforcing `allowNet` host patterns (exact hosts, ports, wildcard subdomains, IPs); the security page says outbound is unrestricted by default while the create page says the default is no outbound access; secrets are placeholders inside the VM and substituted only on requests to approved hosts; HTTP, SSH and VS Code can be exposed",
      "pricing": "Launch post: CPU time $0.05/hour, memory $0.016/GB-hour, volume storage $0.20/GiB-month, with 40 CPU-hours, 1,000 GB-hours and 5 GiB included on the $20/month Pro plan; the same post carries a note dated November 2026 that Deno Sandbox \"is no longer included in Deno Deploy plans and is not available to new customers\"",
      "free_tier": "Free Deno Deploy plans were stopped at their usage limits rather than billed; not available to new customers as of November 2026",
      "compliance": null,
      "regions": "Amsterdam (ams) and Chicago (ord); volumes only in Chicago",
      "best_for": "Short, bursty untrusted-code runs by existing Deno Deploy customers who need egress allow-lists and proxy-injected secrets",
      "overview": [
        "Deno Sandbox is an SDK and API for starting Linux microVMs on Deno Deploy, aimed at running agent-generated or untrusted code. Each sandbox has 2 vCPUs, 768 MB-4 GB of memory and 10 GB of ephemeral disk, boots from a clean image and is torn down when the last reference is dropped or its timeout expires. The distinctive features are an outbound proxy that enforces `allowNet` host allow-lists and a secrets mechanism in which code only ever sees a placeholder, with the real value substituted on requests to approved hosts, as a defence against prompt-injection exfiltration. There are JavaScript (`@deno/sandbox`) and Python (`deno-sandbox`) SDKs, a `deno sandbox` CLI and a REST API.",
        "Persistence comes from read-write volumes and read-only snapshots rather than from pausing a VM; volumes are limited to the Chicago region. The product launched in beta on 3 February 2026 with per-hour CPU and per-GB-hour memory pricing, and the docs still describe a \"pre-release phase\" with a default of 5 concurrent sandboxes per organisation. In November 2026, after Deno announced in October that it was joining Cloudflare, the launch post gained a note that Deno Sandbox is no longer included in Deno Deploy plans and is not available to new customers, and the docs now carry a banner that Deno Deploy will be sunset in April 2027, so this record documents a product that newcomers cannot sign up for."
      ],
      "notable": [
        "Launched in beta on 3 February 2026",
        "Note dated November 2026 on the launch post: no longer included in Deno Deploy plans and not available to new customers; Deno announced on 9 October 2026 that its team is joining Cloudflare",
        "Default concurrency of 5 sandboxes per organisation during the pre-release phase"
      ],
      "pros": [
        "You are an existing Deno Deploy customer who wants microVM code execution inside the same platform",
        "You need egress allow-lists and secrets that never enter the sandbox environment",
        "You want to promote a sandbox into a deployed Deno Deploy app with one call"
      ],
      "cons": [
        "Not available to new customers as of November 2026",
        "Lifetime capped at 30 minutes per the About page; pause/resume of a running VM is not documented",
        "Fixed 2 vCPUs and at most 4 GB memory; volumes only in the Chicago region",
        "Pre-release default of 5 concurrent sandboxes per organisation"
      ],
      "verdict": "A microVM sandbox on Deno Deploy with egress allow-lists and proxy-injected secrets that is not available to new customers as of November 2026, with Deno Deploy due to be sunset in April 2027.",
      "peers": [
        "cloudflare-sandbox",
        "vercel-sandbox",
        "e2b",
        "hopx",
        "fly-sprites"
      ],
      "sources": [
        "https://deno.com/deploy/sandbox",
        "https://docs.deno.com/sandbox/",
        "https://docs.deno.com/sandbox/security/",
        "https://docs.deno.com/sandbox/create/",
        "https://docs.deno.com/sandbox/timeouts/",
        "https://docs.deno.com/sandbox/volumes/",
        "https://deno.com/blog/introducing-deno-sandbox",
        "https://deno.com/deploy/pricing",
        "https://github.com/denoland/sandbox-py",
        "https://deno.com/blog/cloudflare"
      ],
      "status": "winding-down",
      "status_label": "closed to new customers",
      "status_note": "Deno’s launch post carries an update dated November 2026 saying Deno Sandbox is no longer included in Deno Deploy plans and is not available to new customers. Deno announced on October 9, 2026 that its team is joining Cloudflare, and its site says Deno Deploy, which Sandbox runs on, will be sunset in April 2027.",
      "verified_on": "2026-11-14"
    },
    {
      "id": "upstash-box",
      "name": "Upstash Box",
      "company": "Upstash",
      "hq": {
        "region": "California",
        "country": "US",
        "source": "https://upstash.com/about"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://upstash.com/box",
      "docs": "https://upstash.com/docs/box/overall/how-it-works",
      "pricing_url": "https://upstash.com/pricing/box",
      "github": "https://github.com/upstash/box",
      "tagline": "Persistent Docker-container \"boxes\" for coding agents with Claude Code, Codex, OpenCode and Cursor built in, billed on active CPU time rather than wall-clock.",
      "isolation": "Docker container per box with its own filesystem, process tree and network stack; no VM boundary (the product page itself contrasts this with microVM providers)",
      "isolation_short": "Docker container",
      "isolation_class": "container",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Python",
        "CLI",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "No maximum stated (launch post: \"You can keep Upstash Box forever\"); idle boxes auto-pause after 1 hour on the free plan and 6 hours on paid plans, and keep-alive boxes (paid) never pause",
      "max_session_short": "No cap (idle pause 1 h / 6 h)",
      "persistence": "Filesystem (files, packages, git history) persists across pauses and resumes on durable block storage; snapshots capture the filesystem plus agent/runtime settings and can seed many new boxes; memory is not stated to survive a pause",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not mentioned in the docs or pricing",
      "networking": "Outbound allow-all by default (HTTP, HTTPS, DNS, WebSockets, raw TCP); `deny-all` or a `custom` allow-list of domains and CIDRs with denied CIDRs can be set at creation or updated live; private IP ranges and cloud metadata endpoints are always blocked; \"Attach Headers\" inject secret HTTP headers on the host so keys need not enter the box; SSH access by box ID",
      "pricing": "Pay as you go per active CPU hour: Small 2 vCPU/4 GB $0.10, Medium 4 vCPU/8 GB $0.20, Large 8 vCPU/16 GB $0.40; memory not billed; storage $0.10/GB-month with 10 GB included per box; keep-alive (always-on) boxes $8, $16 or $32 per box per month. The pricing page gives two different disk limits for Medium boxes (30 GB in the size table, 20 GB in the FAQ)",
      "free_tier": "Free plan: up to 10 concurrent boxes, 5 CPU hours per month, 2 vCPU/4 GB/10 GB boxes, 1-hour idle timeout",
      "compliance": null,
      "regions": "AWS us-east-1 only, with more regions planned",
      "best_for": "Long-lived per-user coding-agent workspaces where idle time should cost nothing",
      "overview": [
        "Upstash Box, launched on 9 March 2026 by the serverless-data company Upstash, is a \"cloud computer for your agents\": each box is a Docker container with its own filesystem, process tree and network stack, a shell, git, a chosen runtime (Node, Python, Go, Ruby or Rust on Debian or Alpine) and optionally a built-in coding agent (Claude Code, Codex, OpenCode or Cursor, or a custom agent). You drive it from TypeScript or Python SDKs, a CLI, SSH or the hosted Upstash MCP server, sending prompts or commands and receiving structured results.",
        "Boxes are designed to live indefinitely: they pause when idle, keep their files on durable storage and resume on the next request, and snapshots can branch many boxes from one prepared checkpoint. Pricing is the distinctive part: compute is metered on active CPU time rather than wall-clock, so an idle box costs only storage, with fixed-price keep-alive boxes for always-on use. The trade-off is the isolation boundary, which is a container rather than a VM, something the product page acknowledges when comparing itself with microVM-based providers. Only AWS us-east-1 is available today; the Box pages make no compliance statement, and Upstash's general compliance page (SOC 2 Type II) does not name Box."
      ],
      "notable": [
        "Launched 9 March 2026 with a free tier",
        "Billed on active CPU time, not wall-clock: 10% of one core for an hour is quoted at $0.01",
        "Runs on AWS in us-east-1, with more regions planned"
      ],
      "pros": [
        "You want a box an agent returns to for days with files and git history intact",
        "You need idle time to be free, or a flat monthly price for an always-on box",
        "You want Claude Code, Codex, OpenCode or Cursor preconfigured rather than installing an agent yourself",
        "You are already on Upstash and want the same console, API keys and MCP server"
      ],
      "cons": [
        "Container isolation rather than a VM; the product page describes microVMs as \"a stronger boundary than Box's container\"",
        "Single region (AWS us-east-1)",
        "Box is not named on Upstash's compliance page, which states SOC 2 Type II for the platform",
        "Pricing page gives two different Medium disk limits (30 GB and 20 GB)"
      ],
      "verdict": "A persistent workspace for coding agents, billed on active CPU time and isolated by a container rather than a VM.",
      "peers": [
        "boat",
        "exe-dev",
        "fly-sprites",
        "daytona",
        "railway-sandboxes",
        "freestyle"
      ],
      "sources": [
        "https://upstash.com/docs/box/overall/how-it-works",
        "https://upstash.com/docs/box/overall/getstarted",
        "https://upstash.com/docs/box/overall/snapshots",
        "https://upstash.com/docs/box/overall/network-policy",
        "https://upstash.com/pricing/box",
        "https://upstash.com/box",
        "https://upstash.com/blog/upstash-box",
        "https://github.com/upstash/box",
        "https://upstash.com/docs/common/help/compliance"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "tensorlake",
      "name": "Tensorlake Sandboxes",
      "company": "Tensorlake",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.tensorlake.ai/privacy"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.tensorlake.ai",
      "docs": "https://docs.tensorlake.ai",
      "pricing_url": "https://www.tensorlake.ai/pricing",
      "github": "https://github.com/tensorlakeai/tensorlake",
      "tagline": "MicroVM sandboxes with memory snapshots, copy-on-the-fly cloning, warm pools and egress policy, alongside hosted Git and a versioned filesystem, with an Apache-2.0 runtime and BYOC on AWS.",
      "isolation": "MicroVM per sandbox using hardware virtualisation (hosts must expose /dev/kvm); hypervisor not named; GPU sandboxes run on gVisor instead",
      "isolation_short": "MicroVM (gVisor for GPU)",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": true,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Rust",
        "CLI",
        "REST API"
      ],
      "cold_start": "Docs: CPU microVMs \"boot in hundreds of milliseconds\"; homepage cites the vendor's own SQLite benchmark with an 84 ms cold start and about one second to suspend or resume (vendor claims)",
      "cold_start_short": "~100s of ms (claim)",
      "max_session": "Idle timeout (default 600 s), not a wall-clock cap; plan maximums are 2 hours (Free), 24 hours (Usage Credits) and unlimited (Pro, Enterprise); named sandboxes suspend on timeout, ephemeral ones terminate",
      "max_session_short": "2 h / 24 h / unlimited (Pro)",
      "persistence": "Named sandboxes suspend and resume with memory and filesystem intact; filesystem or memory snapshots (checkpoints) persist after termination and can be restored with different resources; `copy` clones a running or suspended sandbox, including memory and processes, into one or many warm-started sandboxes; terminated sandboxes can be restarted under the same ID for 48 hours",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Yes, by request only: H100 GPU sandboxes on gVisor at $3.95 per GPU-hour (reserved), with no live resizing or Tensorlake filesystem mounts",
      "networking": "Outbound open by default; `allow_out` turns the sandbox default-deny (domains, wildcard subdomains, IPv4, CIDRs) and `deny_out` overrides it; policy can be changed at runtime; inbound only to ports listed in `exposed_ports`, authenticated unless explicitly opened; shared NAT egress IPs for allow-listing; SSH over WebSocket PTY and port tunnels",
      "pricing": "Usage Credits / Pro per second: active CPU $0.07 / $0.042 per core-hour (minimum 20% of allocated vCPU), RAM $0.015 / $0.009 per GB-hour, root disk $0.0002 / $0.0001 per GB-hour, snapshot storage $0.07 per GB-month while suspended; egress free. Usage Credits are prepaid packs from $5 (1 CU = $0.01); Pro is $250/month including 25,000 CU with $0.01/CU overage; Enterprise custom",
      "free_tier": "Free plan: 1 concurrent sandbox, 1 vCPU/1 GB/10 GB, sessions up to 2 hours",
      "compliance": "Homepage and docs claim SOC 2 Type II and HIPAA; EU data residency and zero data retention are listed on the docs index",
      "regions": "Not listed; the docs mention EU data residency and BYOC into your own AWS account",
      "best_for": "Agent harnesses, RL rollouts and coding agents that need memory snapshots, cloning and warm pools at scale",
      "overview": [
        "Tensorlake describes itself as \"a serverless runtime for sandboxes and background agentic applications\". Sandboxes are microVMs (hardware virtualisation via KVM; hypervisor unnamed) created from Python, TypeScript or Rust SDKs, the `tl` CLI or HTTP, with SSH, VS Code, Cursor and JetBrains Gateway access. Alongside them sit a versioned POSIX filesystem, managed Git with server-side merges and an orchestration layer for durable workflows. The core runtime is published on GitHub under Apache-2.0, and a self-hosted compute mode runs the executors on bare-metal EC2 in your own AWS account while Tensorlake keeps the control plane.",
        "State handling is the strongest part of the feature set: named sandboxes suspend and resume with memory, checkpoints can be filesystem-only or full memory snapshots, `copy` clones a running sandbox including processes, and pools pre-warm sandboxes. Egress policy is default-open with allow and deny lists that can be changed live. Pricing bills active CPU, allocated RAM and disk per second, with a free plan, prepaid credit packs and a $250/month Pro plan; H100 GPU sandboxes are available by request on gVisor. The homepage lists SOC 2 Type II and HIPAA and names Redpoint and Amplify Partners as backers, and it integrates with Claude Managed Agents, Cursor Cloud Agents, Devin and Harbor for evals and RL."
      ],
      "notable": [
        "Core runtime open source under Apache-2.0 (github.com/tensorlakeai/tensorlake)",
        "Self-hosted compute / BYOC mode runs sandbox executors on i7i.metal EC2 instances in your AWS account",
        "Vendor benchmark claims an 84 ms cold start and a 1.000 score on terminal-bench 2.1 with the oracle agent"
      ],
      "pros": [
        "You need memory snapshots and cloning of running sandboxes for parallel agents or RL rollouts",
        "You want warm pools and queued creation for bursty, high-concurrency workloads",
        "You need to run the sandbox data plane inside your own AWS account",
        "You want hosted Git and a versioned filesystem next to the sandbox"
      ],
      "cons": [
        "Hypervisor is not named in the docs",
        "GPU sandboxes are by request only and use gVisor rather than a VM",
        "No published region list",
        "Free plan allows one 1 vCPU sandbox with 2-hour sessions"
      ],
      "verdict": "A microVM platform for agent harnesses at scale, with memory snapshots, cloning, an Apache-2.0 runtime and a BYOC option on AWS.",
      "peers": [
        "e2b",
        "daytona",
        "modal",
        "morph",
        "blaxel",
        "namespace"
      ],
      "sources": [
        "https://www.tensorlake.ai",
        "https://docs.tensorlake.ai",
        "https://www.tensorlake.ai/pricing",
        "https://docs.tensorlake.ai/sandboxes/lifecycle.md",
        "https://docs.tensorlake.ai/sandboxes/snapshots.md",
        "https://docs.tensorlake.ai/sandboxes/gpu.md",
        "https://docs.tensorlake.ai/sandboxes/quickstart.md",
        "https://docs.tensorlake.ai/sandboxes/networking.md",
        "https://docs.tensorlake.ai/platform/billing.md",
        "https://docs.tensorlake.ai/platform/self-hosted-compute/aws.md",
        "https://docs.tensorlake.ai/llms.txt",
        "https://github.com/tensorlakeai",
        "https://github.com/tensorlakeai/tensorlake",
        "https://docs.tensorlake.ai/sandboxes/introduction.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "namespace",
      "name": "Namespace",
      "company": "Namespace Labs",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://namespace.so",
      "docs": "https://namespace.so/docs",
      "pricing_url": "https://namespace.so/pricing",
      "github": "https://github.com/namespacelabs/integrations",
      "tagline": "Isolated Linux, macOS and Windows VM instances and persistent Devboxes on a developer-compute cloud, with an SDK, egress policy and SOC 2 Type II.",
      "isolation": "Hypervisor-backed VM instances with KVM (nested virtualisation on amd64 Linux); the homepage names a custom hypervisor and scheduler but gives no further detail",
      "isolation_short": "KVM virtual machine",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Go",
        "CLI",
        "MCP server",
        "gRPC/Connect API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Compute instances run to a deadline you set and can extend; no maximum stated. Devboxes auto-shut down on idle (15 min to 8 h presets) with a plan maximum of 3 hours (Developer, Devbox Plus), 5 hours (Team) or 24 hours (Business), raisable to 24 hours via support",
      "max_session_short": "Deadline-based; Devbox 3-24 h",
      "persistence": "Compute instances can be suspended (RAM and disks snapshotted) and woken; volumes and snapshots are managed via the SDK; Devboxes keep persistent storage across stop and start",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not mentioned on the compute, machine-shape or pricing pages",
      "networking": "Egress policies with ALLOW rules in BLOCK or ADVISORY mode, header INJECT of vault secrets and PROXY rules via deep packet inspection, applied to runners, Devboxes and Compute API instances; private networking between instances in a workspace; federation to your own cloud",
      "pricing": "Compute \"unit minutes\" = vCPU (or RAM/2 if higher) x minutes x platform multiplier (Linux 1, Windows 2, Linux on Apple Silicon 7, macOS 10) at $0.001 prepaid or $0.0015 overage per unit minute, 1-minute minimum; e.g. 1 vCPU/2 GB Linux $0.001/min. Plans: Developer $0 pay as you go, Team $100/month (100,000 unit minutes), Business $250/month (250,000), Devbox Plus $250/month (50,000 Devbox Minutes + 400 GB-month). Devboxes $0.004 per Devbox Minute (Linux S=1, M=2, L=4, XL=8 per minute; macOS M=15, L=30) plus $0.20/GB-month storage",
      "free_tier": "30-day free trial of the Developer plan; no card required at signup; no permanent free allowance stated",
      "compliance": "SOC 2 Type II; report and sub-processors via trust.namespace.so; regional preferences on all plans and region-exclusive residency on Enterprise",
      "regions": "Multiple regions worldwide, not individually listed; the demo shows us-east and placement can be constrained by continent or site",
      "best_for": "Teams that want macOS, Windows and Linux VMs for agents and CI from one SDK, plus persistent Devboxes for coding agents",
      "overview": [
        "Namespace is a developer-compute cloud from Namespace Labs, originally built for CI (GitHub Actions runners, Bazel remote execution, Docker builds) and now marketed to AI coding agents with the line \"Sandboxes aren't enough. Your cloud agents need a Devbox.\" Two products matter here. **Compute instances** are hypervisor-isolated Linux (AMD EPYC amd64, Ampere arm64, Apple Silicon in early access), macOS (M4 Pro, M5 Max) and Windows VMs created from a TypeScript or Go SDK, the `nsc` CLI or a gRPC/Connect API, with deadlines, extend, suspend/wake and command execution. **Devboxes** are persistent Linux or macOS development environments with SSH, IDE access and preinstalled agents (Amp, Claude Code, Codex, Copilot, Cursor, OpenCode), managed by a separate CLI and SDK, and used by integrations with Claude Managed Agents, Cursor Cloud Agents and Devin.",
        "Billing is by unit minute with platform multipliers (macOS costs ten times Linux), on a pay-as-you-go Developer plan or $100-$250/month team plans; Devboxes are billed separately per Devbox Minute. Egress policies can allow-list domains, inject secrets as headers and log in advisory mode. Namespace holds SOC 2 Type II and offers regional preferences and enterprise region-exclusive residency. The homepage announced a $42M Series B on 5 October 2026 and names Cursor, ElevenLabs, Ramp, Zed and Warp among more than 1,000 customers."
      ],
      "notable": [
        "$42M Series B announced 5 October 2026 (homepage)",
        "macOS instances up to 16 vCPU/56 GB on Apple M4 Pro and M5 Max, with Xcode 16-27 and simulators",
        "Devbox idle limits of 3 to 24 hours by plan, with 1-50 Devboxes per workspace"
      ],
      "pros": [
        "You need macOS or Windows VMs for agents, not only Linux",
        "You want one platform for CI runners, builds and agent compute with shared caches",
        "You need SOC 2 Type II and data-residency controls",
        "You want persistent Devboxes that agents and humans share over SSH and IDEs"
      ],
      "cons": [
        "Devbox sessions are capped at 3 to 24 hours depending on plan",
        "Billing has a 1-minute minimum and platform multipliers (Windows 2x, macOS 10x Linux)",
        "No cold-start figure, region list or GPU information published",
        "Compute instances and Devboxes are separate products with separate CLIs and billing units"
      ],
      "verdict": "A developer-compute cloud, originally built for CI, whose VMs and Devboxes suit teams that need macOS or Windows instances and SOC 2 alongside Linux.",
      "peers": [
        "depot-sandboxes",
        "ona-gitpod",
        "github-codespaces",
        "coder",
        "railway-sandboxes",
        "tensorlake"
      ],
      "sources": [
        "https://namespace.so",
        "https://namespace.so/docs",
        "https://namespace.so/pricing",
        "https://namespace.so/docs/devbox.md",
        "https://namespace.so/docs/devbox/lifecycle.md",
        "https://namespace.so/docs/platform/compute/linux.md",
        "https://namespace.so/docs/reference/typescript-sdk/compute/manage.md",
        "https://namespace.so/docs/architecture/compute/machine-shapes.md",
        "https://namespace.so/docs/architecture/compute/resource-limits.md",
        "https://namespace.so/docs/workspaces/security.md",
        "https://namespace.so/docs/workspaces/data-residency.md",
        "https://namespace.so/docs/workspaces/billing-and-limits.md",
        "https://namespace.so/docs/security/egress-policy.md",
        "https://namespace.so/docs/reference/sdk.md",
        "https://namespace.so/docs/llms.txt"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "novita-sandbox",
      "name": "Novita Agent Sandbox",
      "company": "Novita AI",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://novita.ai/sandbox",
      "docs": "https://docs.novita.ai/guides/sandbox-overview",
      "pricing_url": "https://docs.novita.ai/guides/sandbox-pricing",
      "github": null,
      "tagline": "Pay-per-second stateful sandboxes from the Novita GPU cloud, with pause/resume, snapshots, templates and secrets, billed by vCPU-second and GiB-second.",
      "isolation": "Not documented beyond \"system-level separation\" and an isolated runtime and filesystem per sandbox; hypervisor or container runtime not named",
      "isolation_short": "Not documented",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "CLI"
      ],
      "cold_start": "Product page: instances launch \"in under 200ms on average\" (vendor claim); resume from pause \"typically takes about one second\"",
      "cold_start_short": "<200 ms (claim)",
      "max_session": "Product page: 1 hour maximum per session on the free tier, 24 hours on the paid tier, custom for Enterprise; the docs describe a creation-time lifetime timeout plus an idle timeout, each of which can pause instead of kill",
      "max_session_short": "1 h free / 24 h paid",
      "persistence": "Pausing preserves filesystem and memory and stops compute billing, with optional auto-resume; snapshots save a sandbox as a reusable environment for new sandboxes (v2 region); templates define starting environments; killed sandboxes keep nothing",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not mentioned for sandboxes; the sandbox docs and pricing list vCPU and RAM only (GPUs are documented as separate Novita products)",
      "networking": "Internet access on by default; can be disabled per sandbox (`allow_internet_access=False`) or restricted with `allow_out` and `deny_out` lists of domains, IPs and CIDRs; any port can be exposed through the sandbox's public HTTPS URL; secrets are injected into HTTPS headers only for allowed hosts",
      "pricing": "Per second: vCPU $0.0000098 per vCPU-second (1-8 vCPU), RAM $0.0000032 per GiB-second (512 MiB to 8 GiB in 512 MiB steps); persistent storage for paused sandboxes, templates and snapshots $0.00009/GB-hour beyond 60 GB free; 20 GB ephemeral disk per running sandbox. Examples: 2 vCPU/1 GiB for 1 hour about $0.08, 8 vCPU/8 GiB for 1 hour about $0.37",
      "free_tier": "Free tier (5 concurrent sandboxes, 2 vCPU/4 GB, 1-hour sessions) until the balance is above $0; $100 promotional credits for new users who complete a setup survey, no card required; 60 GB persistent storage free",
      "compliance": "Footer shows an AICPA SOC 2 badge and links to a trust centre; Enterprise plans list a 99.95% uptime SLA",
      "regions": "us-phx-1 (v2, default; adds Secrets and Snapshots) and us-virginia-1 (v1, legacy); custom region for Enterprise",
      "best_for": "Cheap per-second code execution and browser-agent runs with pause/resume on a single US region",
      "overview": [
        "Novita AI, a GPU cloud and model-inference provider, sells Agent Sandbox as a managed execution environment for AI agents: isolated, stateful sandboxes in which agents run code, install dependencies, drive browsers and keep state across sessions, managed from the `novita-sandbox` Python and Node.js SDKs or a CLI. The docs do not name the isolation technology, saying only that each sandbox is a separate runtime with its own filesystem, processes and network; the product page calls this \"system-level separation\".",
        "Lifecycle is well specified: a lifetime timeout and an idle timeout can each kill or pause the sandbox, pausing keeps memory and filesystem with billing stopped, and the v2 region adds snapshots and encrypted team-scoped secrets. Pricing is per vCPU-second and GiB-second with 60 GB of free persistent storage, which puts an 8 vCPU/8 GiB hour at about $0.37; the free tier caps sessions at one hour and concurrency at five, the paid tier at 24 hours and 100. Two US regions are offered (Phoenix by default, Virginia legacy). Outbound internet access is on by default and can be disabled or restricted with allow and deny lists of domains and CIDRs."
      ],
      "notable": [
        "Paid tier unlocks automatically once the account balance is above $0; $100 promotional credits for new users",
        "v2 region us-phx-1 adds Secrets and Snapshots; us-virginia-1 is a legacy v1 region",
        "Product page claims under 200 ms average launch and lists a 99.95% SLA for Enterprise"
      ],
      "pros": [
        "You want per-second pricing with memory-preserving pause and a generous free storage allowance",
        "You run browser or coding agents that need sessions up to 24 hours",
        "You are already a Novita customer for inference or GPUs"
      ],
      "cons": [
        "Isolation technology is not named in the docs",
        "Two US regions only on standard tiers",
        "Free tier limited to 1-hour sessions and 5 concurrent sandboxes"
      ],
      "verdict": "A per-second sandbox with pause/resume, snapshots and egress allow-lists in two US regions, suited to cost-sensitive workloads; the docs do not name the isolation technology.",
      "peers": [
        "e2b",
        "daytona",
        "hopx",
        "blaxel",
        "cloudflare-sandbox",
        "beam"
      ],
      "sources": [
        "https://novita.ai/sandbox",
        "https://docs.novita.ai/guides/sandbox-overview.md",
        "https://docs.novita.ai/guides/sandbox-learn-more.md",
        "https://docs.novita.ai/guides/sandbox-pricing.md",
        "https://docs.novita.ai/guides/sandbox-your-first-agent-sandbox.md",
        "https://docs.novita.ai/guides/sandbox-internet-access.md",
        "https://docs.novita.ai/guides/sandbox-faq-isolation.md",
        "https://docs.novita.ai/guides/sandbox-faq-regions.md",
        "https://docs.novita.ai/guides/sandbox-quota-limit.md",
        "https://docs.novita.ai/llms.txt"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "hopx",
      "name": "HopX",
      "company": "Bunnyshell",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://hopx.ai",
      "docs": "https://docs.hopx.ai",
      "pricing_url": "https://hopx.ai/pricing/",
      "github": "https://github.com/hopx-ai",
      "tagline": "Firecracker microVM sandboxes from Bunnyshell with template-based sizing, memory-preserving pause, VNC desktop automation, an MCP server and $200 of free credits.",
      "isolation": "Firecracker Linux microVM per sandbox with hardware-level kernel isolation (homepage); the docs overview says only \"isolated computing environments\"",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": "Homepage: about 100 ms startup from prebuilt snapshots (vendor claim)",
      "cold_start_short": "~100 ms (claim)",
      "max_session": "No expiry unless `timeout_seconds` is set; the homepage says sandboxes can run for hours, days or weeks; `set_timeout` resets the timer from now and the sandbox is destroyed, data included, on expiry",
      "max_session_short": "No cap",
      "persistence": "Pause preserves full state including memory, variables and running processes and resumes where it left off; stop keeps the filesystem only; templates package environments and sandboxes launch from prebuilt snapshots; no user-created snapshot API documented",
      "snapshots": null,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not mentioned in docs or pricing",
      "networking": "Per-sandbox `internet_access` flag (false creates an isolated sandbox); default not stated; the security page lists \"optional egress controls per sandbox\", but the docs do not describe an allow-list; exposed ports get region-scoped URLs",
      "pricing": "Pay per second with no plans, minimums or idle charges: vCPU $0.000014/s, memory $0.0000045/s per GiB, storage $0.00000003/s per GiB (about $0.050 per vCPU-hour and $0.016 per GiB-hour); examples on the page: 8 hours at 2 vCPU/4 GB about $1.50; enterprise custom SLAs and volume discounts via sales",
      "free_tier": "$200 free credits, no card required",
      "compliance": "Security page lists SOC 2 (independent assessment), ISO/IEC 27001 and ISO 9001:2015 (vendor claims); certificates or bridge letters available under NDA on request",
      "regions": "Not listed; the API exposes a region field for filtering and URLs but no region names are documented",
      "best_for": "Pay-per-second microVM sandboxes with desktop automation and no session limits",
      "overview": [
        "HopX (\"HopX by Bunnyshell\", from the environments-as-a-service company Bunnyshell) provides sandboxes for running untrusted code, scripts, tests, jobs and AI agents. The homepage describes them as Firecracker Linux microVMs that start in about 100 ms from prebuilt snapshots. Access is through Python (`hopx-ai`) and JavaScript (`@hopx-ai/sdk`) SDKs, a CLI, a REST API and an MCP server for IDEs and assistants; capabilities include synchronous and streaming code execution in Python, JavaScript, Bash and Go, file operations, background processes, environment variables and desktop automation over VNC with mouse, keyboard and screenshots.",
        "Resources are fixed by template (vCPU 1-64, memory 512 MB-64 GB, disk 1-250 GB depending on plan), sandboxes have no expiry unless a timeout is set, and pause preserves memory and processes while stop keeps only the filesystem. Pricing is purely per second with no tiers and $200 of free credits. The docs do not cover regions, egress allow-lists or snapshots of user state; the separate security page lists SOC 2, ISO/IEC 27001 and ISO 9001:2015 and mentions optional per-sandbox egress controls, and the terms of service name Bunnyshell Inc. as the contracting party."
      ],
      "notable": [
        "$200 free credits with no card required",
        "Marketed as \"HopX by Bunnyshell\"; status page at status.bunnyshell.com",
        "Templates allow up to 64 vCPU, 64 GB memory and 250 GB disk, plan-dependent"
      ],
      "pros": [
        "You want Firecracker isolation with pay-per-second pricing and no monthly plan",
        "You need sandboxes that never expire unless you say so",
        "You want desktop automation (VNC, mouse, keyboard, screenshots) inside the sandbox",
        "You want an MCP server so an IDE assistant can open sandboxes directly"
      ],
      "cons": [
        "Region names are not published, and egress controls are mentioned on the security page but not described in the docs",
        "Resources cannot be set per sandbox, only via templates",
        "No user-created snapshot or fork API documented"
      ],
      "verdict": "A per-second Firecracker sandbox with desktop tooling and no session cap, from Bunnyshell; region names are not published.",
      "peers": [
        "e2b",
        "daytona",
        "novita-sandbox",
        "blaxel",
        "tensorlake",
        "scrapybara"
      ],
      "sources": [
        "https://hopx.ai",
        "https://hopx.ai/pricing/",
        "https://docs.hopx.ai",
        "https://docs.hopx.ai/llms.txt",
        "https://docs.hopx.ai/core-concepts/sandboxes/creating.md",
        "https://docs.hopx.ai/core-concepts/sandboxes/timeout.md",
        "https://docs.hopx.ai/core-concepts/sandboxes/managing-state.md",
        "https://docs.hopx.ai/core-concepts/templates/configuration.md",
        "https://github.com/hopx-ai",
        "https://hopx.ai/security",
        "https://hopx.ai/terms"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "baseten-sandboxes",
      "name": "Baseten Sandboxes",
      "company": "Baseten",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.baseten.co/terms-and-conditions/"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.baseten.co",
      "docs": "https://docs.baseten.co/sandboxes/overview",
      "pricing_url": "https://www.baseten.co/pricing/",
      "github": "https://github.com/basetenlabs/baseten-go",
      "tagline": "Standby-and-resume sandboxes on the Baseten inference platform, based on primitives from Blaxel, which Baseten acquired, with Python, TypeScript and Go SDKs, volumes and usage-based pricing.",
      "isolation": "MicroVM per sandbox with its own kernel and hardware-level isolation (launch post, 5 November 2026); the docs say only \"an isolated environment\" and the hypervisor is not named. Baseten acquired Blaxel in 2026 and said it would add products based on Blaxel's primitives, starting with Sandboxes",
      "isolation_short": "MicroVM, own kernel",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Sandboxes expire 7 days after creation by default; expiry policies are TTL_MAX_AGE, TTL_IDLE and a fixed DATE, one of each type, with no maximum stated",
      "max_session_short": "7 days default",
      "persistence": "Standby keeps files and pauses running processes until the next request; files are deleted on expiry or delete; persistent volumes (grow-only, one sandbox at a time, region-bound) keep files across sandboxes; custom images are versioned per team; no user-created snapshot API is documented, although the launch post says standby snapshots the filesystem, memory and running processes",
      "snapshots": null,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not mentioned on the sandbox configure page; vCPU and memory only",
      "networking": "HTTPS previews expose a web app in a sandbox via public or private URLs; short-lived sandbox access tokens for requests and MCP connections; egress controls are not yet documented for sandboxes (the launch post says an Agent Proxy for outbound control will come online \"in the coming months\")",
      "pricing": "Pricing page: CPU $0.0655 per vCPU-hour, memory $0.0211 per GB-hour, snapshot storage $0.20/GB-month, volume storage $0.16/GB-month, stored images $0.045/GB-month; at launch (5 November 2026) the docs described the product as \"free to use until November 16\"",
      "free_tier": null,
      "compliance": null,
      "regions": "us-pdx-1 and us-was-1 appear in the docs; default is the closest region",
      "best_for": "Coding and research agents on Baseten that need a persistent workspace with standby and volumes next to hosted models and RL training",
      "overview": [
        "Baseten, an inference and training platform, launched Sandboxes on 5 November 2026 after acquiring Blaxel in September 2026. A sandbox is an isolated environment where an agent runs commands, background processes and file operations, created from built-in or custom images with 1-32 vCPUs and 1-8 GiB of memory per vCPU (default 2 vCPU/4 GiB). You manage it from Python (`baseten`), TypeScript (`@basetenlabs/sandbox`) or Go SDKs, the `baseten sandbox` CLI (marked pre-release), a management REST API, MCP clients or the dashboard. The docs do not name the isolation technology, but the launch post says each sandbox gets its own microVM with its own kernel; Baseten said it would add products based on Blaxel's primitives, starting with Sandboxes.",
        "Sandboxes go on standby when they stop receiving requests, keeping files and pausing processes, and resume on the next request; they expire after 7 days by default under configurable max-age, idle or date policies. Persistent volumes carry files between sandboxes within a region. The pricing page lists $0.0655 per vCPU-hour and $0.0211 per GB of RAM per hour, and at launch the docs described the product as free to use until 16 November 2026, with 10 concurrent sandboxes per organisation. The docs also include worked examples that train models with RL against sandboxed coding tasks, which is the platform's natural pairing."
      ],
      "notable": [
        "Based on primitives from Blaxel; Baseten announced on 10 September 2026 that it had acquired Blaxel",
        "At launch the docs described the product as free to use until 16 November 2026, with a limit of 10 concurrent sandboxes per organisation, raisable via support",
        "CLI sandbox commands are labelled pre-release"
      ],
      "pros": [
        "You run inference or RL training on Baseten and want agent sandboxes in the same account",
        "You want standby-and-resume with persistent volumes rather than ephemeral runs",
        "You need Go as well as Python and TypeScript SDKs"
      ],
      "cons": [
        "Hypervisor is not named; the microVM isolation is described in the launch post, not in the docs",
        "Concurrency limited to 10 per organisation by default",
        "Egress controls, a user-facing snapshot API and compliance statements are not documented for sandboxes"
      ],
      "verdict": "A microVM sandbox inside the Baseten platform, launched in November 2026 on primitives from the acquired Blaxel, that suits teams already running inference or training on Baseten.",
      "peers": [
        "blaxel",
        "modal",
        "beam",
        "koyeb",
        "northflank",
        "fly-sprites"
      ],
      "sources": [
        "https://docs.baseten.co/sandboxes/overview.md",
        "https://docs.baseten.co/sandboxes/manage.md",
        "https://docs.baseten.co/sandboxes/configure.md",
        "https://docs.baseten.co/sandboxes/lifecycle.md",
        "https://docs.baseten.co/sandboxes/volumes.md",
        "https://docs.baseten.co/_llms/documentation.txt",
        "https://www.baseten.co/blog/blaxel-is-joining-baseten-to-build-the-future-of-agentic-cloud/",
        "https://www.baseten.co/blog/building-the-agent-cloud-introducing-baseten-sandboxes",
        "https://www.baseten.co/pricing/"
      ],
      "verified_on": "2026-11-15"
    },
    {
      "id": "railway-sandboxes",
      "name": "Railway Sandboxes",
      "company": "Railway",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://railway.com/sandboxes",
      "docs": "https://docs.railway.com/sandboxes",
      "pricing_url": "https://railway.com/sandboxes",
      "github": "https://github.com/railwayapp/railway-ts-sdk",
      "tagline": "On-demand Linux VMs inside a Railway environment with forks, checkpoints, Docker, private networking to your services and per-second billing; generally available since September 2026.",
      "isolation": "Isolated Linux virtual machine per sandbox; hypervisor not named",
      "isolation_short": "Linux VM",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "CLI",
        "GraphQL API"
      ],
      "cold_start": "Product page: \"ready in seconds\"; no figure",
      "cold_start_short": null,
      "max_session": "No maximum lifetime stated; idle timeout destroys the sandbox after 30 minutes by default on Hobby and Pro (1-120 minutes, or disabled) and after 5 minutes on Free and Trial (1-5 minutes, cannot be disabled)",
      "max_session_short": "No cap (idle 30 min default)",
      "persistence": "Sandboxes are ephemeral; checkpoints are named server-side disk snapshots that outlive the sandbox and boot where captured; forks copy a running sandbox's disk (not memory or processes); templates cache build steps; pause/resume is not documented",
      "snapshots": true,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not mentioned in the sandbox docs",
      "networking": "Outbound internet via NAT on all sandboxes; ISOLATED (default) has no access to environment services, PRIVATE joins the environment's private network to reach databases by hostname; up to 10 public HTTPS domains per sandbox set at creation; SSH and port forwarding with an account SSH key; no egress allow-list documented",
      "pricing": "Per second at Railway VM rates: $0.00001929 per GB-second of memory in use (about $50/GB-month), $0.00001929 per vCPU-second of active CPU (about $50/vCPU-month), egress $0.05/GB; idle sandboxes still bill memory; usage draws from the plan's included credit. Example from the page: 20 minutes at 1 GB and 30% of a vCPU about $0.03",
      "free_tier": "Available on Free and Trial plans within included usage: 10 sandboxes per environment at 2 vCPU/2 GB with a 5-minute idle timeout",
      "compliance": null,
      "regions": "US West (default), US East, EU West and Southeast Asia; forks stay in the source region",
      "best_for": "Coding agents and test runs that need to reach your Railway databases and services over the private network",
      "overview": [
        "Railway Sandboxes are isolated Linux VMs that Railway, the application-hosting platform, creates on demand inside one of your project environments. The Ubuntu base image ships git, Node and the Claude Code, Codex, OpenCode and Pi agents, and every sandbox includes a Docker daemon. You drive them from the open-source `railway` TypeScript SDK, the Railway CLI, the dashboard or the GraphQL API, with a files API, durable sessions that keep 8 MiB of output for reattachment, SSH and port forwarding. The feature graduated from Priority Boarding to general availability on 17 September 2026 and is available on every plan.",
        "The differentiator is environment scoping: a sandbox can join the environment's private network and talk to your Postgres or other services by hostname, or stay isolated with outbound-only internet. State is handled with forks (disk copies of running sandboxes), named checkpoints and cached templates, but pause/resume is not documented and idle sandboxes are destroyed after a configurable timeout. Billing is per second at Railway's standard memory and active-CPU rates, drawn from the plan's included usage; plan caps run from 10 sandboxes at 2 vCPU on Free to 100 at up to 32 vCPU on Pro. The hypervisor is not named."
      ],
      "notable": [
        "Generally available since 17 September 2026, after graduating from Priority Boarding",
        "Four regions: US West, US East, EU West and Southeast Asia",
        "Every sandbox includes the Docker daemon and CLI"
      ],
      "pros": [
        "You run services on Railway and want agents to test against them over the private network",
        "You need forks and checkpoints to try approaches in parallel",
        "You want Docker inside the sandbox without configuration",
        "You want EU and Asian regions as well as US"
      ],
      "cons": [
        "Pause/resume is not documented; idle sandboxes are destroyed, and memory continues to bill while idle",
        "Hypervisor and compliance are not documented on the sandbox pages",
        "An egress allow-list is not documented; every sandbox has outbound internet access",
        "Free and Trial sandboxes are limited to 2 vCPU/2 GB and a 5-minute idle timeout"
      ],
      "verdict": "On-demand Linux VMs inside a Railway environment with private networking to your existing services; suited to teams already deploying on Railway.",
      "peers": [
        "vercel-sandbox",
        "fly-sprites",
        "koyeb",
        "northflank",
        "modal",
        "depot-sandboxes"
      ],
      "sources": [
        "https://docs.railway.com/sandboxes.md",
        "https://docs.railway.com/sandboxes/quickstart",
        "https://railway.com/sandboxes.md",
        "https://railway.com/changelog/2026-09-18-railway-sandboxes.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "docker-cloud-sandboxes",
      "name": "Docker Cloud Sandboxes",
      "company": "Docker",
      "hq": {
        "city": "Palo Alto",
        "region": "California",
        "country": "US",
        "source": "https://www.docker.com/company/contact/"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.docker.com/products/docker-sandboxes/",
      "docs": "https://docs.docker.com/ai/sandboxes/cloud/",
      "pricing_url": "https://docs.docker.com/agentic-platform/signup/",
      "github": null,
      "tagline": "The cloud counterpart to Docker's local agent sandboxes: the same `sbx` CLI and Kits running microVMs on Docker-managed compute, metered per second under the experimental Docker Agentic Platform.",
      "isolation": "MicroVM with a separate kernel per sandbox and its own Docker Engine, \"the same microVM isolation and policies as local sandboxes\"; hypervisor not named",
      "isolation_short": "MicroVM, own kernel",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": null,
      "sdks": [
        "CLI",
        "JavaScript/TypeScript",
        "REST API"
      ],
      "cold_start": "Press release: boots in the \"low hundreds of milliseconds\" with secrets, policy and agent configuration preconfigured (vendor claim)",
      "cold_start_short": "~100s of ms (claim)",
      "max_session": "Expires after 1 hour by default; the timeout action can stop a resumable sandbox or delete it, and the press release says agents can run unattended for hours; no maximum stated",
      "max_session_short": "1 h default TTL",
      "persistence": "Sandboxes that can be resumed are stopped rather than deleted at expiry; `sbx move` copies a filesystem snapshot between cloud and local (processes and memory are not transferred); the API manages images, snapshots and volumes; files are deleted on `rm`",
      "snapshots": true,
      "pause_resume": true,
      "gpu": false,
      "gpu_detail": "No - cloud sandboxes have no host hardware access (GPU, USB, display are local-only)",
      "networking": "Deny-by-default policy at account and sandbox level, opened per host with `--allow-network` (e.g. github.com:443); credentials are injected by a proxy so raw values never enter the VM; ports are exposed through public HTTPS URLs; L7 method/path filtering is not supported in the cloud; the security page (written for local sandboxes) says sandboxes cannot communicate directly over the network",
      "pricing": "Pay-as-you-go compute metered in seconds by runtime and CPU/memory configuration under a Docker Agentic Platform subscription with no recurring fee; rates are not published; 1-16 vCPUs per the press release; model-provider inference billed separately",
      "free_tier": "None stated for cloud; the `sbx` CLI and local sandboxes are free",
      "compliance": null,
      "regions": null,
      "best_for": "Developers already using Docker Sandboxes locally who want the same policies and Kits when a run must outlive the laptop",
      "overview": [
        "Docker announced Cloud Sandboxes on 24 September 2026 as the cloud extension of its local Docker Sandboxes product. The same `sbx` CLI, with a global `--cloud` flag, creates, attaches to, copies files into and removes sandboxes that run as microVMs on Docker-managed infrastructure instead of the developer's machine, so agentic runs continue after a laptop closes and teams can scale without their own servers. Kits (OCI images packaging an agent, tools and access rules) and policies work identically in both places, and `sbx move` carries a filesystem snapshot between local and cloud. A Sandboxes REST API with an OpenAPI spec and a TypeScript SDK let applications create cloud sandboxes directly.",
        "Cloud and local sandboxes keep separate state, credentials, secrets, network policy and volumes; cloud sandboxes cannot mount host paths or use host hardware, expose ports via public HTTPS URLs, and expire after one hour by default. Access requires the Docker Agentic Platform subscription, which the docs label experimental, with compute billed per second at unpublished rates and no recurring fee. Regions, compliance and size tables are not published on the pages read. Because the site already lists the local product, this record covers only what is cloud-specific."
      ],
      "notable": [
        "Announced 24 September 2026 at WeAreDevelopers North America, \"available today\"",
        "Requires `sbx` CLI 0.45.0 or later and a Docker Agentic Platform subscription, which the docs call experimental",
        "Docker plans to submit the Kits spec to the CNCF"
      ],
      "pros": [
        "You already use Docker Sandboxes locally and want identical policies and Kits in the cloud",
        "You need agent runs to continue after your laptop shuts down",
        "You want credential injection and deny-by-default networking enforced outside the VM"
      ],
      "cons": [
        "Compute rates, regions and compliance are not published",
        "The Docker Agentic Platform subscription is labelled experimental",
        "No host hardware (GPU, display, USB) and no L7 network filtering in the cloud",
        "Cloud and local state do not interoperate beyond filesystem moves"
      ],
      "verdict": "The cloud counterpart to Docker Sandboxes for existing users, offered under the Docker Agentic Platform, which Docker labels experimental and for which per-unit compute rates are not published.",
      "peers": [
        "docker-sandboxes",
        "e2b",
        "daytona",
        "depot-sandboxes",
        "railway-sandboxes",
        "vercel-sandbox"
      ],
      "sources": [
        "https://www.docker.com/press-release/cloud-sandboxes-extending-secure-ai-agent-isolation-beyond-the-laptop/",
        "https://docs.docker.com/ai/sandboxes/",
        "https://docs.docker.com/ai/sandboxes/cloud/",
        "https://docs.docker.com/ai/sandboxes/cloud/local-vs-cloud/",
        "https://docs.docker.com/ai/sandboxes-api/",
        "https://docs.docker.com/ai/sandboxes/architecture/",
        "https://docs.docker.com/ai/sandboxes/security/",
        "https://docs.docker.com/agentic-platform/signup/",
        "https://www.docker.com/products/docker-sandboxes/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "depot-sandboxes",
      "name": "Depot Sandboxes",
      "company": "Depot",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://depot.dev",
      "docs": "https://depot.dev/docs/sandboxes/overview",
      "pricing_url": "https://depot.dev/pricing",
      "github": null,
      "tagline": "MicroVM sandboxes on the bare-metal compute behind Depot CI, with a TypeScript SDK, Tailscale networking and a Managed Agents service that hosts Claude Code, Codex, Devin and Amp sessions.",
      "isolation": "MicroVM per sandbox on bare-metal hosts (\"Real microVMs, not containers\") with a full Linux syscall interface, Docker and nested VMs; hypervisor not named",
      "isolation_short": "MicroVM on bare metal",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": true,
      "sdks": [
        "JavaScript/TypeScript",
        "CLI"
      ],
      "cold_start": "Blog: \"fast boot from a pre-cached base image\"; no figure",
      "cold_start_short": null,
      "max_session": "Runs until stopped or the configured timeout expires; the SDK launch post gives a default timeout of two hours with extensions; no maximum stated; a stopped sandbox cannot be restarted",
      "max_session_short": "2 h default timeout",
      "persistence": "Snapshots save the home and /workspace directories (not processes or memory) and seed new sandboxes; Managed Agents snapshot the workspace automatically between runs, excluding node_modules, .venv and target; stopped sandboxes cannot be resumed; persistent disk was listed as planned",
      "snapshots": true,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": "Not mentioned",
      "networking": "Sandboxes can join your organisation's Tailscale tailnet to reach private services; Depot Cache credentials are injected by default; run secrets are scoped to up to 16 hosts each; no egress allow-list documented",
      "pricing": "Developer $20/month with 2,000 included minutes, Startup $200/month with 20,000, Business custom (dedicated infrastructure, deployment in your own AWS account); minutes are shared across Depot CI, Sandboxes and Managed Agents and measured at 2 vCPU, so the default 4 vCPU sandbox uses 2 minutes per minute; overage $0.00005 per vCPU-second, billed per second with no minimum; storage 25 GB / 250 GB included then $0.20/GB-month",
      "free_tier": "7-day trial on Developer and Startup plans; a permanent free tier is not listed",
      "compliance": null,
      "regions": null,
      "best_for": "Teams on Depot CI who want agent sandboxes and hosted coding-agent sessions billed from the same minute pool",
      "overview": [
        "Depot, the container-build and CI acceleration company, exposes the compute behind Depot CI as sandboxes: microVMs on bare-metal hosts with NVMe storage, on x86-64 or ARM64, with Docker and nested virtualisation. Two surfaces exist. The **Sandbox SDK** (`@depot/sandbox`, Node.js, announced as a private beta on 18 June 2026) creates sandboxes, streams commands and exposes a `node:fs/promises`-style filesystem. **Managed Agents** hosts sessions from Amp, Claude Code, Claude Managed Agents, Devin and OpenAI in Depot sandboxes of 2-16 vCPUs, replacing the earlier Claude Code remote agent product, with automatic workspace snapshots, Tailscale access to private services and a `depot-mcp` bridge.",
        "Snapshots capture files only and stopped sandboxes cannot restart, so this is an ephemeral-run model rather than a persistent computer. Pricing is per second against a monthly minute pool shared with Depot CI and Managed Agents; the Business tier offers deployment in your own AWS account. The hypervisor, regions and compliance are not documented on the sandbox pages, and the SDK's current status beyond the June private-beta note is not stated."
      ],
      "notable": [
        "Sandbox SDK announced in private beta on 18 June 2026; Managed Agents replaced the earlier Claude Code remote agent product",
        "Sizes from 2 to 64 vCPUs on Intel or Arm; Managed Agents at 2, 4, 8 or 16 vCPUs",
        "Business plan offers dedicated infrastructure deployed in your own AWS account"
      ],
      "pros": [
        "You use Depot CI and want sandbox and agent minutes from the same pool with shared caches",
        "You want hosted Claude Code, Codex, Devin or Amp sessions that survive closing your laptop",
        "You need sandboxes on your Tailscale tailnet to reach private services",
        "You want Arm as well as x86 microVMs"
      ],
      "cons": [
        "Snapshots save files only, not processes or memory; a stopped sandbox cannot be restarted",
        "TypeScript-only SDK, and it was private beta at announcement",
        "Hypervisor, regions and compliance not documented",
        "No free plan is listed; Developer and Startup start with a 7-day trial"
      ],
      "verdict": "A microVM sandbox and hosted-agent service that shares a minute pool with Depot CI, with file-only snapshots and a Node.js SDK, aimed at existing Depot customers.",
      "peers": [
        "namespace",
        "railway-sandboxes",
        "docker-cloud-sandboxes",
        "e2b",
        "blaxel",
        "modal"
      ],
      "sources": [
        "https://depot.dev/docs/sandboxes/overview",
        "https://depot.dev/docs/agents/overview",
        "https://depot.dev/pricing",
        "https://depot.dev/changelog/2026-06-18-sandbox-sdk-beta.md",
        "https://depot.dev/blog/now-available-the-depot-sandbox-sdk"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "exe-dev",
      "name": "exe.dev",
      "company": "exe.dev",
      "hq": {
        "city": "Berkeley",
        "region": "California",
        "country": "US",
        "source": "https://exe.dev/docs/terms-of-service"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://exe.dev",
      "docs": "https://exe.dev/docs",
      "pricing_url": "https://exe.dev/pricing",
      "github": null,
      "tagline": "Persistent root VMs for people and their agents, driven entirely over SSH (or the same commands via HTTPS and MCP), with replicated NVMe storage, VM copies and flat-rate CPU/RAM pools in nine regions.",
      "isolation": "Each VM runs as its own VMM process inside a shared CPU/RAM pool; hypervisor not named",
      "isolation_short": "VM (VMM per VM)",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": true,
      "sdks": [
        "CLI (over SSH)",
        "HTTPS API",
        "MCP server"
      ],
      "cold_start": "API explorer: creating a VM takes \"about a second\" (vendor claim)",
      "cold_start_short": "~1 s (claim)",
      "max_session": "VMs are long-lived and persistent with no session limit; stopped standalone VMs keep their disk",
      "max_session_short": "No cap",
      "persistence": "Persistent VM disks on local NVMe, streamed to other machines in the regional cluster for durability; `cp` copies a VM's disk and config; S3 backups (development preview) with incremental 1 MB chunks and 7-day minimum retention; no memory snapshots documented",
      "snapshots": true,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not mentioned",
      "networking": "Egress `all` by default or `integrations` (blocks the public internet, keeps metadata and attached integrations); public or private HTTPS app URLs via an SNI/Host-routed TLS proxy and global load balancer; SSH with public-key auth only; an LLM gateway adds provider keys outside the VM",
      "pricing": "Personal: $15/month for a 2 vCPU/4 GiB pool, $35 (4/8), $75 (8/16), $155 (16/32), up to 50 VMs, 100 GiB disk, 200 GiB bandwidth. Work: $75/month per team ($37.50 for the first 3 months) including 2 vCPU/4 GiB, extra capacity $0.105 per 2 vCPU/4 GiB unit-hour, pools of 2-512 vCPUs, 1,000 VMs per pool. Standalone VMs metered per second at the same unit rate ($0.0525/h for 1 vCPU/2 GiB to $0.84/h for 16 vCPU/32 GiB). Enterprise custom with dedicated hardware and AWS/GCP VPC",
      "free_tier": "7-day trial of the Personal plan, no card required when linking a paid ChatGPT subscription; no permanent free tier",
      "compliance": null,
      "regions": "Los Angeles, New York, Dallas, Frankfurt, Tokyo, Sydney, Singapore, London (Oregon closed to new accounts); the account \"lobby\" server is in the United States",
      "best_for": "Developers who want a fleet of persistent root VMs they and their agents reach over SSH, at a flat monthly price",
      "overview": [
        "exe.dev, from a small Bay Area team, is \"lots of computers\": persistent Linux VMs with root that you create and manage by SSHing to exe.dev (`ssh exe.dev new`), with the same commands available through a single HTTPS endpoint and a hosted MCP server so any agent can be given its own VMs. Each VM is a separate VMM process in a pool of CPU and RAM that your VMs share, with disks on local NVMe replicated within the region. Features include public or private HTTPS URLs for apps on a VM, VM copying, a two-mode egress policy, an LLM gateway that injects provider keys, Google Calendar and Gmail connections for agents, and S3-backed incremental backups in preview.",
        "Pricing is a flat monthly pool rather than per sandbox: Personal plans from $15/month for 2 vCPU/4 GiB shared by up to 50 VMs, Work plans at $75/month per team with metered extra pool capacity, and standalone VMs with reserved resources at $0.105 per 2 vCPU/4 GiB hour. Nine regions span the US, Europe and Asia-Pacific, and Enterprise adds single-tenant hardware or BYOC into AWS or GCP. The hypervisor, compliance certifications and any SDKs beyond the SSH/HTTPS commands are not documented."
      ],
      "notable": [
        "Nine regions (US, Frankfurt, London, Tokyo, Sydney, Singapore), with the Oregon region closed to new accounts",
        "Personal plan from $15/month for a 2 vCPU/4 GiB pool shared by up to 50 VMs",
        "VM backups to S3 are a development preview"
      ],
      "pros": [
        "You want many small persistent VMs for a flat monthly price rather than per-second metering",
        "You and your agents work over SSH and want the same commands via HTTPS or MCP",
        "You need regions in Asia-Pacific and Europe as well as the US",
        "You want app URLs and an LLM key gateway without running a proxy yourself"
      ],
      "cons": [
        "Hypervisor and compliance are not documented",
        "Language SDKs are not documented; the API is one POST endpoint wrapping SSH commands with a 30-second timeout and no stdin",
        "Egress control is a two-mode switch (`all` or `integrations`); a host allow-list is not documented",
        "Memory snapshots and pause/resume are not documented"
      ],
      "verdict": "A flat-rate fleet of persistent SSH-driven VMs for agent-assisted developers, rather than an SDK-driven sandbox API for high-volume automation.",
      "peers": [
        "boat",
        "upstash-box",
        "fly-sprites",
        "freestyle",
        "railway-sandboxes"
      ],
      "sources": [
        "https://exe.dev",
        "https://exe.dev/docs",
        "https://exe.dev/pricing",
        "https://exe.dev/security",
        "https://exe.dev/about",
        "https://exe.dev/docs/architecture",
        "https://exe.dev/docs/infrastructure",
        "https://exe.dev/docs/https-api",
        "https://exe.dev/docs/regions",
        "https://exe.dev/docs/vm-backups",
        "https://exe.dev/docs/internet-access",
        "https://exe.dev/docs/billing/overview",
        "https://exe.dev/docs/standalone-vms",
        "https://exe.dev/docs/mcp"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "pullrun",
      "name": "Pullrun",
      "company": "Pullrun",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://pullrun.dev",
      "docs": "https://github.com/pullrun/pullrun#readme",
      "pricing_url": null,
      "github": "https://github.com/pullrun/pullrun",
      "tagline": "Apache-2.0 sandbox runtime that runs the same OCI image as a runc container, a Firecracker microVM on Linux or an Apple Virtualization VM on macOS, with a CLI and MCP server.",
      "isolation": "Firecracker microVM with a per-VM kernel on KVM (Linux), Apple Virtualization.framework VM (macOS) or runc container, chosen per workload",
      "isolation_short": "Firecracker / Apple VM / runc",
      "isolation_class": "mixed",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "CLI",
        "MCP server"
      ],
      "cold_start": "README (vendor benchmark on an Apple M3, cold cache): Firecracker cold boot ~500 ms, warm ~200 ms; Apple Virtualization VM boot ~160 ms; container run ~400 ms",
      "cold_start_short": "~500 ms (claim)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "`commit` writes a running workload's filesystem to a new image layer; stopped VMs retain state (Firecracker and Apple VM rootfs persist, container rootfs is ephemeral); `save`/`load` export tarballs",
      "snapshots": true,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Five `--net` modes: isolated, bridge, slirp, host and none, with IPAM, DNS and iptables integration; per-pod bridge subnets with NAT under the Kubernetes CRI shim",
      "pricing": "Free to self-host under Apache-2.0. The website describes enterprise \"per-node pricing\" and a planned hosted API \"billed per active sandbox-second\", with no figures published",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Running agent workloads from OCI images as microVMs on your own Linux hosts or Apple Silicon Macs",
      "overview": [
        "Pullrun is an open-source runtime, written in Rust and Go, that takes an OCI image and runs it either as a runc container, as a Firecracker microVM with its own kernel on Linux (KVM) or as an Apple Virtualization.framework VM on Apple Silicon Macs; Windows is supported through WSL2. It ships as a single binary with a 30-plus-command CLI, Docker Compose compatibility, a Kubernetes CRI shim (labelled beta) and a built-in MCP server exposing 15 tools for agent clients such as Claude Code, Cursor and opencode. Images are stored in a content-addressed store so many concurrent sandboxes share one copy of each layer.",
        "State is handled at the image level: `commit` turns a running workload's filesystem into a new layer and stopped VMs keep their rootfs, but there is no documented pause/resume of a running VM. Security controls include Cosign signature verification, SBOM checks, seccomp profiles, read-only rootfs and secrets encrypted at rest. The project is at v0.7.x; the website calls the container runtime stable, the CRI shim beta and the Firecracker and Apple VM backends \"feature-complete but evolving\". A hosted sandbox API and managed control plane are described as planned, so today it is self-host only."
      ],
      "notable": [
        "Version 0.7.8–0.7.9 at verification, 134 GitHub stars, Apache-2.0 with a CLA for contributions.",
        "README: single 12 MB binary with a 24.6 MiB idle daemon RSS; Firecracker cold boot ~500 ms and Apple VM boot ~160 ms (vendor benchmark on an Apple M3).",
        "Hosted API and managed control plane are listed as \"Planned\"; the Kubernetes CRI shim is labelled beta."
      ],
      "pros": [
        "You want one runtime that boots the same OCI image as a container or a microVM on Linux and on Apple Silicon.",
        "You need an Apache-2.0 sandbox layer you can self-host with no central daemon.",
        "You run agents through MCP clients and want a built-in MCP server rather than an SDK.",
        "You want image-level commits and air-gapped save/load of sandbox state."
      ],
      "cons": [
        "No hosted service yet; the managed API is only \"planned\".",
        "Language SDKs are not documented; interaction is through the CLI, Compose files or MCP.",
        "Pause/resume of running VMs, GPU support and session limits are not documented.",
        "At v0.7.x; the website describes the Firecracker and Apple VM backends as \"feature-complete but evolving\"."
      ],
      "verdict": "A self-hosted OCI-to-microVM runtime for teams who want to own the sandbox layer; a hosted API is listed as planned.",
      "peers": [
        "microsandbox",
        "firecracker",
        "boxlite",
        "smolvm",
        "docker-sandboxes",
        "apple-container"
      ],
      "sources": [
        "https://pullrun.dev",
        "https://github.com/pullrun/pullrun"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "superserve",
      "name": "Superserve",
      "company": "Superserve",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.superserve.ai/privacy"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.superserve.ai",
      "docs": "https://docs.superserve.ai",
      "pricing_url": "https://www.superserve.ai/pricing/",
      "github": "https://github.com/superserve-ai/superserve",
      "tagline": "Persistent Firecracker microVM sandboxes for long-running agents, with pause/resume, memory-and-disk snapshots, forks and a credentials broker that keeps secrets out of the sandbox.",
      "isolation": "Firecracker microVM per sandbox",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": true,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": "Docs: sandboxes spin up \"in seconds\"; no figure published",
      "cold_start_short": null,
      "max_session": "Unlimited runtime on Developer and Pro plans; optional auto-pause cap on active time and an idle timeout of 60 s to 7 days",
      "max_session_short": "Unlimited",
      "persistence": "Pause snapshots memory and disk and suspends the VM; paused sandboxes are kept indefinitely by default (optional auto-delete up to 30 days); snapshots of memory and disk can be forked into new sandboxes and persist until deleted",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Egress allow/deny rules by CIDR or domain; per-sandbox network log of HTTP/HTTPS connections; preview URLs with public or per-port authenticated access; secrets are bound to a sandbox and injected by a proxy on outbound calls so credentials never enter the VM",
      "pricing": "Usage on every plan: $0.0504 per vCPU-hour and $0.0162 per GiB-hour, billed per second while running; storage $0.000108 per GiB-hour for retained data including paused sandboxes. Developer $20/month (includes $20 usage, 100 concurrent, 4 vCPU/16 GiB max); Pro $250/month (includes $250 usage, 2,000 concurrent, 16 vCPU/64 GiB max); Enterprise custom",
      "free_tier": "Limited free trial for a first sandbox, no card required (amount not stated)",
      "compliance": "Trust centre hosted on Vanta; no certifications named",
      "regions": "Each team has a fixed home region; available regions are not listed",
      "best_for": "Long-running coding and personal agents that need a persistent VM with pause, forks and brokered credentials",
      "overview": [
        "Superserve is a hosted sandbox service in which every sandbox is a Firecracker microVM that can stay alive indefinitely, pause with its memory and disk intact and be snapshotted or forked. You drive it from Python or TypeScript SDKs, a CLI, a REST API or a hosted MCP server, and the docs carry integration guides for Claude Code, Codex, OpenAI Agents SDK, LangChain, CrewAI and others, plus a `superserve/desktop` template that ships an XFCE desktop and Chrome for computer-use agents. Templates are built from a BuildSpec, and Docker can run inside a sandbox.",
        "The distinctive piece is the secrets layer: credentials are bound to a sandbox and injected by a proxy on outbound requests, with an audit trail of which credential was used where. Pricing is per-second vCPU and memory plus retained storage, with Developer ($20/month) and Pro ($250/month) plans and an Enterprise tier that can deploy into your own AWS, GCP or Azure VPC for early partners. The SDKs, CLI and console are Apache-2.0 on GitHub; source for the sandbox service itself is not published. No maturity label, compliance certification or region list is published."
      ],
      "notable": [
        "Snapshots: 100 per team, 20 per sandbox; a running sandbox pauses \"usually for under a second\" while one is taken.",
        "Paused sandboxes are kept forever by default; idle timeout is configurable from 60 seconds to 7 days.",
        "Client repository (SDKs, CLI, MCP server, console) is Apache-2.0 with 512 GitHub stars at verification."
      ],
      "pros": [
        "You want sandboxes with no runtime limit that pause and resume with memory intact.",
        "You need API credentials kept outside the sandbox with an audit log of their use.",
        "You run coding-agent harnesses and want ready-made integration guides and a desktop template.",
        "You want per-second billing with no idle charge for paused sandboxes."
      ],
      "cons": [
        "No cold-start figure, region list or compliance certification is published.",
        "Only the client SDKs, CLI and console are published as open source; source for the sandbox runtime is not published.",
        "BYOC deployment is limited to Enterprise \"early partners\"."
      ],
      "verdict": "A well-documented persistent-VM sandbox whose credentials broker suits agents that must call third-party APIs safely.",
      "peers": [
        "e2b",
        "daytona",
        "morph",
        "withruntime",
        "leap0",
        "instavm"
      ],
      "sources": [
        "https://www.superserve.ai",
        "https://www.superserve.ai/pricing/",
        "https://docs.superserve.ai/llms.txt",
        "https://docs.superserve.ai/introduction.md",
        "https://docs.superserve.ai/sandbox/lifecycle.md",
        "https://docs.superserve.ai/sandbox/snapshots.md",
        "https://docs.superserve.ai/sandbox/desktop.md",
        "https://github.com/superserve-ai/superserve",
        "https://app.vanta.com/superserve.ai/trust/6ouhg1woewexrlytvw99e"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "opencomputer",
      "name": "OpenComputer",
      "company": "OpenComputer (Digger)",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://opencomputer.dev/contact"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://opencomputer.dev",
      "docs": "https://docs.opencomputer.dev",
      "pricing_url": "https://opencomputer.dev",
      "github": "https://github.com/diggerhq/opencomputer",
      "tagline": "Persistent cloud Linux VMs for long-running agents that hibernate and wake, with checkpoints, regions and a serverless-agent harness on top.",
      "isolation": "Full Linux VM per sandbox; the self-hosting guide names QEMU/KVM, the website does not name the hypervisor. The repository README, which now describes a coding agent used from Slack, says each agent session is isolated with rootless Podman on a VM",
      "isolation_short": "QEMU/KVM VM",
      "isolation_class": "vm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Python",
        "CLI",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "No limit documented; VMs are described as always on, with hibernate and wake",
      "max_session_short": "No limit documented",
      "persistence": "VMs keep state across sessions and can hibernate and wake; \"instant snapshots\" with fork or restore to any point; CPU and memory resizable at runtime; files under /workspace are kept in a versioned S3-backed workspace",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Open internet by default; an `[egress]` allowlist of up to 64 domains (HTTPS only) and 64 IPs/CIDRs can be set per agent; fixed outbound IPs per region; preview URLs with Clerk auth and custom domains",
      "pricing": "Free: shared up to 2 vCPU and 2 GB, 3 GB disk, 5 concurrency. Medium $300/month: 8 CPU, 16 GB, 50 GB disk, 50 concurrency. Large $3,000/month: 64 CPU, 512 GB, 3,800 GB disk, 500 concurrency. Custom: contact",
      "free_tier": "Free tier: shared up to 2 vCPU and 2 GB memory, 3 GB disk, 5 concurrency",
      "compliance": null,
      "regions": "Selectable per agent: broad `us`, `eu`, `ap` plus sub-regions such as us-east, eu-west, uk, jp, au, ca, mx, me, sa and af; default is wherever capacity is available, usually the US",
      "best_for": "B2B agent platforms that give each user or agent a persistent Linux VM rather than a disposable sandbox",
      "overview": [
        "OpenComputer, which its contact page says is \"built by a small team at Digger\", sells persistent cloud Linux VMs for agents and positions them against stateless sandboxes: each machine is a full VM with its own filesystem and OS access that stays on, hibernates when idle and wakes on demand. On top of the VMs sits a \"serverless agents\" harness in which agents are TypeScript functions deployed with the `opencomputer` CLI, with durable sessions, schedules, webhooks, cross-session memory, a per-project SQL database and Slack, GitHub and Linear connections. There are TypeScript and Python SDKs and a Management API.",
        "Checkpoints can be forked or restored, VMs can be resized at runtime and sandboxes can be pinned to one of about twenty regions. Pricing is by machine pool rather than per second: a free shared tier, Medium at $300/month and Large at $3,000/month. The repository is Apache-2.0 and includes a SELFHOSTING.md that describes an \"operator path\" running worker VMs on QEMU/KVM with PostgreSQL, Redis and S3-compatible storage on AWS or Azure, rather than a one-command install. Hibernate/wake timings, GPU support and compliance are not documented. As of November 2026 the repository README and the docs introduction describe OpenComputer as a coding agent that a team works with in Slack, running on your own VM or the hosted service with each session isolated with rootless Podman; the website continues to describe persistent cloud VMs for agents."
      ],
      "notable": [
        "As of November 2026 the repository README describes OpenComputer as \"The open-source Devin alternative: a coding agent your team works with in Slack\", and the docs introduction opens \"OpenComputer turns any VM into a multiplayer coding agent\"; the website still reads \"Persistent cloud VMs for AI agents\".",
        "Repository has 629 GitHub stars and is Apache-2.0; self-hosting is described as \"an operator path, not a one-command hosted product\".",
        "Blog post \"Scaling OpenComputer from one VM to a million sandboxes\" (June 2026) mentions per-second billing from 10-second heartbeats; the plans listed on the website are monthly machine pools.",
        "Comparison pages target coding-agent products (Claude Code on the web, Codex cloud, Devin) rather than sandbox SDKs."
      ],
      "pros": [
        "You want a VM that keeps state for days and hibernates rather than a sandbox that expires.",
        "You need region pinning with fixed egress IPs for allowlisting on the other end.",
        "You are building an agent product and want sessions, schedules, webhooks and memory handled for you.",
        "You want an Apache-2.0 codebase you could run yourself with enough operations effort."
      ],
      "cons": [
        "The website does not name the hypervisor; QEMU/KVM appears only in the self-hosting guide.",
        "No hibernate/wake timing, GPU or compliance information is published.",
        "Published plans are monthly machine pools ($300 and $3,000 a month above the free tier) rather than per-second rates.",
        "The documentation is organised around OpenComputer's own agent harness (Slack agent, sessions API and Serverless Agents)."
      ],
      "verdict": "Persistent Linux VMs for agents that hibernate and wake, sold as monthly machine pools with a built-in agent harness; suited to B2B agent platforms that give each user or agent its own machine.",
      "peers": [
        "exe-dev",
        "fly-sprites",
        "superserve",
        "islo",
        "daytona"
      ],
      "sources": [
        "https://opencomputer.dev",
        "https://docs.opencomputer.dev/llms.txt",
        "https://docs.opencomputer.dev/agents/regions.md",
        "https://docs.opencomputer.dev/agents/egress.md",
        "https://docs.opencomputer.dev/agents/sessions.md",
        "https://github.com/diggerhq/opencomputer",
        "https://raw.githubusercontent.com/diggerhq/opencomputer/main/SELFHOSTING.md",
        "https://opencomputer.dev/contact",
        "https://raw.githubusercontent.com/diggerhq/opencomputer/main/README.md",
        "https://docs.opencomputer.dev/",
        "https://docs.opencomputer.dev/getting-started/introduction.md"
      ],
      "verified_on": "2026-11-15"
    },
    {
      "id": "baponi",
      "name": "Baponi",
      "company": "Baponi",
      "hq": {
        "city": "Austin",
        "region": "Texas",
        "country": "US",
        "source": "https://baponi.ai/about"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://baponi.ai",
      "docs": "https://baponi.ai/docs",
      "pricing_url": "https://baponi.ai/pricing",
      "github": null,
      "tagline": "Call-per-execution code sandbox built on a hardened nsjail, with sub-20 ms setup, persistent thread state and bring-your-own-bucket storage.",
      "isolation": "Hardened nsjail (Linux namespaces, seccomp-bpf syscall filtering, cgroups v2, zero capabilities) sharing the host kernel; no hypervisor",
      "isolation_short": "nsjail (OS-level)",
      "isolation_class": "os",
      "open_source": false,
      "license": null,
      "self_host": true,
      "byoc": true,
      "sdks": [
        "Python",
        "REST API",
        "MCP server"
      ],
      "cold_start": "Docs: \"typically 12–18ms for sandbox setup\"; website claims \"<20ms\" cold-start overhead on every call",
      "cold_start_short": "<20 ms (claim)",
      "max_session": "Per execution: 60 seconds on Free, 1 hour on Pro, unlimited on Enterprise (timeout range 1–3600 s); nothing runs between calls",
      "max_session_short": "60 s / 1 h (Pro)",
      "persistence": "With a `thread_id`, the diff of /home/baponi is saved to cloud storage after each call and restored on the next (retention 24 hours, 7 days, 30 days or forever); /data mounts (your own S3, GCS or Azure bucket, or managed volumes) are always persistent; /tmp is ephemeral",
      "snapshots": null,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Blocked by default (no network interface, no DNS); `unrestricted` mode allows outbound internet with private IPs and cloud metadata endpoints always blocked, 200 Mbps per core; API keys can only downgrade a policy",
      "pricing": "Credits: 1 credit = 60 seconds at 1 CPU/1 GiB, scaling with resources. Free $0 with 1,000 credits/month; Pro $97/month with 10,000 credits and $1.00 per 1,000 extra, up to 4 CPU/4 GiB, 100 concurrent runs; Enterprise custom (self-hosted). Extra managed storage $0.10/GB/month",
      "free_tier": "Free plan: 1,000 credits/month, 5 concurrent runs, 60 s max execution, 1 CPU/1 GiB, 10 GB managed storage, no card required",
      "compliance": "No certification is named as held (\"Formal certifications in progress\"); trust centre lists SOC 2 Type 1 preparation \"kickoff Q4 2026\" and a DPA available for review",
      "regions": null,
      "best_for": "High-volume, short code executions from agents and MCP clients where state is kept in your own bucket",
      "overview": [
        "Baponi is a hosted code-execution service whose tagline is \"Run code, not containers\": each API call spins up a fresh isolated environment, runs the code and tears it down, so there is no container lifecycle to manage and no idle billing. Isolation is OS-level, built on a hardened build of Google's nsjail with namespaces, seccomp-bpf, cgroups v2 and zero capabilities, with the orchestration layer written in Rust; Baponi says it contributes to nsjail. Python 3.14, Node.js 25 and Bash are available by default and custom OCI images can be imported. Access is via a REST endpoint, a Python SDK or an MCP server for Claude Desktop, Cursor and similar clients, with connectors for databases, Git hosts and Slack.",
        "State persists through a `thread_id`, which saves the home-directory diff between calls, and through `/data` mounts that can point at your own S3, GCS or Azure bucket (read-only and sub-path scoping supported) or Baponi-managed volumes. Pricing is credit-based rather than per second: Free with 1,000 credits a month, Pro at $97/month, and an Enterprise tier that deploys the whole platform into your Kubernetes cluster or VPC with a Helm chart, including air-gapped installs. SOC 2 Type 1 preparation is scheduled for Q4 2026; no certification is held yet."
      ],
      "notable": [
        "Docs: sandbox setup \"typically 12–18ms\"; the website advertises \"<20ms\" on every call.",
        "Trust centre (dated June 2026) lists SOC 2 Type 1 preparation starting Q4 2026 and a penetration test due within 90 days of first production deployment.",
        "Enterprise self-hosting uses a single Helm chart on any Kubernetes cluster, with bring-your-own OIDC, PostgreSQL and storage."
      ],
      "pros": [
        "You want very fast per-call execution with no sandbox lifecycle to manage.",
        "You need the sandbox to read and write your own cloud bucket as a local directory.",
        "You run agents through MCP clients and want a hosted MCP endpoint.",
        "You are an enterprise that must self-host on Kubernetes, including air-gapped."
      ],
      "cons": [
        "OS-level isolation (nsjail) with no hypervisor or guest kernel; a microVM option is not documented.",
        "Executions are capped at 60 s on Free and 1 hour on Pro; no long-lived processes.",
        "No compliance certification yet; SOC 2 Type 1 work starts Q4 2026.",
        "Only a Python SDK is documented; a TypeScript SDK is not."
      ],
      "verdict": "A fast, stateless-by-default execution API with bucket-backed state, best for interpreter-style bursts rather than long-lived agent workspaces.",
      "peers": [
        "e2b",
        "codapi",
        "judge0",
        "hopx",
        "modal",
        "upstash-box"
      ],
      "sources": [
        "https://baponi.ai",
        "https://baponi.ai/docs.md",
        "https://baponi.ai/docs/getting-started/how-it-works.md",
        "https://baponi.ai/pricing",
        "https://baponi.ai/trust"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "islo",
      "name": "Islo",
      "company": "Incredibuild",
      "hq": {
        "city": "Tel Aviv",
        "country": "IL",
        "source": "https://www.incredibuild.com/contact-us"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://islo.dev",
      "docs": "https://docs.islo.dev/overview",
      "pricing_url": "https://islo.dev/pricing/",
      "github": null,
      "tagline": "Persistent microVM \"Machines\" for autonomous coding agents from the Incredibuild team, with a credentials gateway, snapshots and lifecycle policies, billed per vCPU-hour.",
      "isolation": "MicroVM with its own kernel per agent computer; hypervisor not named",
      "isolation_short": "MicroVM (hypervisor not named)",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": true,
      "sdks": [
        "CLI",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "No fixed limit documented; lifecycle policies set `pause_after_idle`, `pause_after` (total active runtime) and `delete_after` (from creation)",
      "max_session_short": "Policy-driven",
      "persistence": "Pausing snapshots the VM and releases compute; files, processes and in-memory state return on resume (auto-resume on activity); snapshots can be saved, listed and used to create new sandboxes",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Gateway profiles enforce access policies and inject credentials at the gateway boundary so agents never hold keys; every outbound request is logged, allowed or denied; port forwards; SSH, scp and rsync through an Islo proxy",
      "pricing": "$0.07 per vCPU-hour, $0.04 per GB-hour of memory and $0.0007 per GB-hour of disk, billed only while a Machine runs (example: 2 vCPU and 4 GB for 2 hours about $0.60); no seat licences or idle fees; per-task cost limits; enterprise and BYOC terms by contact",
      "free_tier": "Sign-up credits with no card required (amount not stated)",
      "compliance": "Not stated; compliance documentation by request to support@islo.dev",
      "regions": "Not listed; the API exposes an endpoint listing the compute regions allowed for a tenant",
      "best_for": "Engineering teams running autonomous coding agents on full repositories with brokered access to GitHub, Slack and Linear",
      "overview": [
        "Islo is built by the team behind Incredibuild and describes itself as \"infra for safely running autonomous coding agents in real, secure dev environments\". Its unit is a Machine: a prepared environment with repositories, dependencies and tools that runs in a microVM with its own kernel, stays up for hours and keeps state between runs. Around the Machines sits a factory model of Agents, Lines (repeatable workflows triggered manually, on a schedule or by GitHub, Slack, Linear or Sentry events) and Gateways that inject credentials and log every outbound call. You work through the `islo` CLI (`islo use`, `islo snapshot`, `islo ssh`), a REST API with an OpenAPI 3.1 spec, or an MCP server for the docs.",
        "Machines pause, resume and delete under lifecycle policies, and snapshots of warm environments restore without rebuilding. Billing is per vCPU-hour, GB-hour and GB-hour of disk with no seat fee; the control plane and Machines can also run inside your own VPC. Under a \"trusted by\" heading the site shows logos including Microsoft, Amazon, Adobe and HSBC and a reinforcement-learning solution for Harbor benchmarks and rollouts. The hypervisor, machine size limits, startup times, regions and compliance status are not documented, and no Python or TypeScript SDK is listed."
      ],
      "notable": [
        "\"Built by the team behind Incredibuild\"; footer carries \"© 2026 Incredibuild\".",
        "Pricing: $0.07 per vCPU-hour, $0.04 per GB-hour memory, $0.0007 per GB-hour disk, billed while running.",
        "API reference lists roughly 260 endpoints including sandboxes, snapshots, gateway profiles, cloud roles and compute regions."
      ],
      "pros": [
        "You want persistent, pausable microVM dev environments rather than disposable sandboxes.",
        "You need agents to use GitHub, Slack or Linear without ever holding the credentials.",
        "You run scheduled or event-triggered agent workflows across a codebase.",
        "You want BYOC deployment of the control plane and machines into your own cloud account."
      ],
      "cons": [
        "Hypervisor is not named and no startup, size or region figures are published.",
        "No Python or TypeScript SDK is documented; the CLI and REST API are the documented interfaces.",
        "Compliance certifications are not stated on the site.",
        "Machines sit inside Islo's own model of Factories, Agents and Lines."
      ],
      "verdict": "An agent-computer platform built by the team behind Incredibuild, suited to teams who want workflow orchestration and credential brokering alongside the sandbox.",
      "peers": [
        "ona-gitpod",
        "daytona",
        "superserve",
        "opencomputer",
        "coder",
        "namespace"
      ],
      "sources": [
        "https://islo.dev",
        "https://islo.dev/machines/",
        "https://islo.dev/security/",
        "https://islo.dev/pricing/",
        "https://docs.islo.dev/overview",
        "https://docs.islo.dev/llms.txt",
        "https://docs.islo.dev/concepts/sandbox-lifecycle",
        "https://docs.islo.dev/cli/sandbox-commands.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "leap0",
      "name": "Leap0",
      "company": "Leap0",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://leap0.dev",
      "docs": "https://leap0.dev/docs",
      "pricing_url": "https://leap0.dev",
      "github": "https://github.com/leap0-dev",
      "tagline": "Firecracker microVM sandboxes in public preview, free during preview, with memory snapshots, forks, a per-sandbox firewall with credential-injecting header transforms and an XFCE desktop.",
      "isolation": "Firecracker microVM with a dedicated Linux kernel per sandbox; the VMM process is jailed with chroot, cgroup v2, seccomp filters and unique UID/GID pairs",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": true,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "MCP server"
      ],
      "cold_start": "Website: \"Sandboxes boot in around 100ms\" (vendor claim)",
      "cold_start_short": "~100 ms (claim)",
      "max_session": "Sandbox timeout from 1 minute to 8 hours, default 5 minutes; limits may change during the public preview",
      "max_session_short": "8 h max",
      "persistence": "Point-in-time snapshots of a running sandbox including in-memory state and writable disk; pause, snapshot and resume across sessions; fork one snapshot into many parallel sandboxes; 10 GiB disk per sandbox",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Per-sandbox firewall with domain allowlists, CIDR rules and header transforms that inject secrets outside the sandbox; demo shows outbound TCP blocked and DNS restricted by default; HTTP/SSE/WebSocket access to apps inside the sandbox",
      "pricing": "Free during public preview. Rates to apply when billing goes live: $0.0504 per vCPU-hour ($0.000014 per second) and $0.0162 per GB-hour ($0.0000045 per GB-second)",
      "free_tier": "Free during public preview, no credit card required",
      "compliance": null,
      "regions": "United States only; more regions planned",
      "best_for": "Trying a full-featured microVM sandbox (desktop, LSP, git, snapshots) at no cost while it is in public preview",
      "overview": [
        "Leap0 is a hosted sandbox service in public preview. Each sandbox is a Firecracker microVM with its own kernel, with the VMM jailed by chroot, cgroup v2 and seccomp. Templates can be any container image from public registries or private ECR, Artifact Registry or ACR. The feature set is broad for a preview product: Python, TypeScript and Go SDKs plus an MCP server; a code interpreter; filesystem, git, process and PTY APIs; Python and TypeScript language servers; SSH and VNC; and a full XFCE desktop at 1440x900 with screenshots, mouse, keyboard and screen recording for computer-use agents.",
        "Snapshots capture memory and writable disk and can be forked into many sandboxes; sandboxes run 1–8 vCPU, 512 MiB–8 GB and 10 GiB disk with a timeout of up to 8 hours. Networking is locked down by a per-sandbox firewall whose header transforms let secrets be injected outside the sandbox. The service is free during the preview, with per-second vCPU and memory rates published for when billing starts; the preview page warns that breaking changes and short downtime may occur. Sandboxes run only in the US, no compliance information is published, and the vendor offers BYOC and on-premises deployment through sales."
      ],
      "notable": [
        "Public preview: free, no card, with the docs stating \"Breaking changes may happen\" and \"Short periods of downtime may occur\".",
        "Limits: 1–8 vCPU, 512 MiB–8 GB memory, 10 GiB disk, timeout 1 minute to 8 hours, 1,000 API requests per 60 seconds.",
        "Published future rates: $0.000014 per vCPU-second and $0.0000045 per GB-second."
      ],
      "pros": [
        "You want to evaluate a microVM sandbox with desktop, LSP and git support at no cost.",
        "You need memory-inclusive snapshots and forks for parallel exploration.",
        "You want outbound secrets injected by the firewall rather than placed in the sandbox.",
        "You run computer-use agents and want a ready XFCE desktop with recording."
      ],
      "cons": [
        "Public preview with explicit warnings of breaking changes and downtime; billing not yet live.",
        "US region only; sessions capped at 8 hours.",
        "Compliance information and the operating company's legal name are not published on the site."
      ],
      "verdict": "A Firecracker microVM sandbox service in public preview, free during the preview, whose docs state that breaking changes and short downtime may occur; suited to evaluating its desktop, snapshot and fork features.",
      "peers": [
        "e2b",
        "instavm",
        "superserve",
        "withruntime",
        "morph",
        "scrapybara"
      ],
      "sources": [
        "https://leap0.dev",
        "https://leap0.dev/docs",
        "https://leap0.dev/docs/limits",
        "https://leap0.dev/docs/public-preview",
        "https://github.com/leap0-dev"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "qbox",
      "name": "qbox",
      "company": "qbox (Utibeabasi Umanah)",
      "hq": null,
      "kind": "software",
      "tier": 2,
      "alternatives": true,
      "website": "https://qbox.sh",
      "docs": "https://qbox.sh/docs",
      "pricing_url": null,
      "github": null,
      "tagline": "Self-hosted Firecracker microVM orchestrator (alpha) for agents and untrusted code, installed with one script on any Linux host with KVM.",
      "isolation": "Firecracker microVM on KVM with its own kernel per sandbox",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": null,
      "license": null,
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Python",
        "CLI",
        "REST API"
      ],
      "cold_start": "Website: sandboxes resume from template snapshots \"in well under a second\" (claim, no figure)",
      "cold_start_short": "<1 s (claim)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Templates are baked from OCI images or Git repos into warm snapshots that sandboxes resume from; Python and JavaScript kernel contexts can be reused across calls; object-storage-backed volumes mount at a path and can be shared across sandboxes",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Control plane and guest talk over a private channel with no public SSH or inbound exposure; guest ports can be exposed as preview URLs; outbound access can be scoped or cut off; activity feed streams network, DNS and TLS events",
      "pricing": "Not published; contact hi@qbox.sh",
      "free_tier": null,
      "compliance": "None listed; \"no telemetry, no phone-home\" and no data leaves your perimeter",
      "regions": "Your own hosts",
      "best_for": "Operators who want Firecracker sandboxes on their own bare metal or KVM-capable VMs without Kubernetes",
      "overview": [
        "qbox is a self-hosted sandbox platform by Utibeabasi Umanah, labelled alpha, that runs untrusted and AI-generated code in Firecracker microVMs on hardware you operate. A single installer (`curl -fsSL https://qbox.sh/install.sh | sh`) checks for KVM, TUN and Docker, stages Firecracker and jailer, downloads a guest kernel and the `qboxd` guest agent, and starts a Compose stack of PostgreSQL, NATS, object storage, a control plane, dashboard, template builder, host agent and Caddy for TLS. Hosts need Linux x86_64 with `/dev/kvm` (bare metal, `.metal` instances or nested virtualisation); arm64 is not supported yet.",
        "Templates from OCI images or Git repos are baked into warm snapshots, sandboxes expose a Python SDK, a `qbox` CLI and a REST plus WebSocket API, and code can run in persistent Python or JavaScript kernels. Headless Chromium sessions can be driven over CDP and a WebRTC live view is included. The site stresses that nothing phones home. The docs do not publish a licence or a source repository: the installer pulls container images from a GHCR account, so whether the code is open source is undocumented. There is no admission gate yet, so overloading a host degrades rather than queues."
      ],
      "notable": [
        "Labelled \"alpha\" on the site; requirements page says \"There's no admission gate yet\" and arm64 \"isn't supported yet\".",
        "Minimum host: Linux x86_64 with /dev/kvm, 2 vCPU, 4 GB RAM, 20 GB disk; installer pins Firecracker v1.12.1.",
        "No pricing, licence or source repository is published; contact is hi@qbox.sh."
      ],
      "pros": [
        "You want Firecracker isolation on your own hosts with a one-script install and no Kubernetes.",
        "You need sandboxes that never phone home, for regulated or air-gapped environments.",
        "You want persistent Python/JavaScript kernels and headless Chromium in the same sandbox."
      ],
      "cons": [
        "Labelled alpha; the docs say there is no admission gate yet, so spawning past host capacity degrades and may fail.",
        "Licence, source availability and pricing are not documented.",
        "x86_64 Linux only; no hosted option.",
        "Pause/resume and GPU support are not documented."
      ],
      "verdict": "A self-hosted Firecracker microVM orchestrator, labelled alpha and with no published licence, for operators who want sandboxes on their own KVM-capable Linux hosts.",
      "peers": [
        "microsandbox",
        "arrakis",
        "firecracker",
        "opensandbox",
        "pullrun"
      ],
      "sources": [
        "https://qbox.sh",
        "https://qbox.sh/docs",
        "https://qbox.sh/docs/self-hosting",
        "https://qbox.sh/docs/requirements",
        "https://qbox.sh/install.sh",
        "https://qbox.sh/docs/api-reference"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "instavm",
      "name": "InstaVM",
      "company": "InstaVM",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://instavm.io",
      "docs": "https://instavm.io/docs",
      "pricing_url": "https://instavm.io/pricing",
      "github": "https://github.com/instavm",
      "tagline": "Firecracker-based Linux microVMs for agents with sub-200 ms claimed boots, snapshots, clones, persistent volumes, a desktop runtime and SSH access, with a local Apple-container companion (CodeRunner).",
      "isolation": "Firecracker microVM with its own kernel, filesystem and network stack per sandbox",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "CLI",
        "REST API"
      ],
      "cold_start": "Website: \"185ms\" P95 cold start, \"Provision a fresh VM in ~147ms\" over SSH, snapshot restore \"typically under 500ms\" (vendor claims)",
      "cold_start_short": "<200 ms (claim)",
      "max_session": "Session lifetime is configurable; no maximum documented; always-on VMs are offered",
      "max_session_short": "Configurable",
      "persistence": "State persists across `execute()` calls within a session and is destroyed when it ends; named volumes persist across VM lifecycles with mid-task checkpoints; VMs can be paused, snapshotted, resumed or cloned; snapshots built from OCI images",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Default-deny outbound with domain/CIDR allowlists and package-manager controls (Free plan egress limited to known package managers and AI APIs; Pro unrestricted or allowlists); proxy-based secret injection; shares expose any port publicly or privately with custom domains",
      "pricing": "Per second: $0.000014 per vCPU (about $0.05/hour) and $0.0000045 per GB of RAM (about $0.02/hour); 10 GB volume storage included, then $0.0002 per GB-hour. Free plan $0 base plus usage (5 concurrent VMs); Pro $100/month plus usage (100 concurrent VMs, 99.9% SLA); Enterprise custom",
      "free_tier": "$50 in starting credits with a card on file; 5 concurrent VMs and 100 browser sessions a month",
      "compliance": "\"SOC 2 Type II compliance\" listed as an Enterprise plan feature; no report details published",
      "regions": null,
      "best_for": "Agent builders who want fast-booting microVMs with desktop, SSH and snapshot-clone workflows under per-second billing",
      "overview": [
        "InstaVM is a hosted platform that boots a full Linux microVM per sandbox, based on Firecracker, with a dedicated kernel, filesystem and network stack. You reach it through Python and TypeScript SDKs, an `instavm` CLI, a REST API or plain `ssh instavm.dev`, and an agent skill exists for Claude Code, Codex, Gemini CLI and others. Sandboxes default to 2 vCPU and 512 MB (1–8 vCPU, 256–8,192 MB) and run Python 3.11+ and Node.js 18+ with common tools. A desktop runtime with browser, terminal, sudo and noVNC live takeover supports computer-use agents, and `instavm deploy` (marked experimental) can publish an app from a VM with a share URL.",
        "Persistence comes from named volumes with checkpoints, pause/snapshot/resume/clone of VMs and always-on VMs for deployed apps or MCP servers. Billing is per second for vCPU and RAM with no idle charges, on a Free plan ($50 starting credit) or Pro at $100/month plus usage; Enterprise adds dedicated infrastructure and SOC 2 Type II. Outbound traffic is default-deny with allowlists. The companion CodeRunner project (Apache-2.0, 901 GitHub stars) runs agent code locally on Apple Silicon inside Apple's `container` VMs. Regions, GPU support and maximum session length are not documented."
      ],
      "notable": [
        "Website: \"185ms\" P95 cold start and \"Provision a fresh VM in ~147ms\" (vendor figures).",
        "Site badge claims \"#1 on Hacker News\" with 1,143 votes; the Enterprise plan lists SOC 2 Type II and a 99.99% SLA.",
        "CodeRunner, the local Mac companion, is Apache-2.0 with 901 GitHub stars; Linux support via Firecracker is on its roadmap."
      ],
      "pros": [
        "You want fast microVM boots with per-second billing and a $50 trial.",
        "You need desktop and browser automation with live noVNC takeover.",
        "You want to clone a running VM or checkpoint a volume mid-task.",
        "You prefer SSH-first access without installing an SDK."
      ],
      "cons": [
        "Free plan lists egress to known package managers and AI APIs; unrestricted egress is listed under Pro.",
        "Regions, GPU support and maximum session length are not documented.",
        "SOC 2 Type II is listed only as an Enterprise feature with no report details.",
        "A card must be on file even for the free tier."
      ],
      "verdict": "A Firecracker-based microVM platform for agent builders with per-second billing, desktop and SSH access; regions and maximum session length are not documented.",
      "peers": [
        "e2b",
        "leap0",
        "superserve",
        "withruntime",
        "scrapybara",
        "cua"
      ],
      "sources": [
        "https://instavm.io",
        "https://instavm.io/pricing",
        "https://instavm.io/docs",
        "https://instavm.io/docs/quickstart",
        "https://instavm.io/docs/concepts/sandboxes",
        "https://github.com/instavm/coderunner"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "tenki-sandbox",
      "name": "Tenki Sandbox",
      "company": "Tenki Cloud (Luxor Technology)",
      "hq": {
        "city": "Seattle",
        "region": "Washington",
        "country": "US",
        "source": "https://tenki.cloud/company/about.md"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://tenki.cloud/products/sandbox",
      "docs": "https://tenki.cloud/docs/sandbox/quickstart",
      "pricing_url": "https://tenki.cloud/pricing",
      "github": "https://github.com/LuxorLabs",
      "tagline": "Full Linux VMs for agents from the Tenki CI-runner company, booting from Ceph-backed disks in under a second (claimed), with pause, memory snapshots, forks and volumes at per-second rates.",
      "isolation": "Hardware-isolated single-tenant VM with its own kernel; hypervisor not named; disks mount from a Ceph cluster",
      "isolation_short": "Single-tenant VM",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Python",
        "Go",
        "CLI"
      ],
      "cold_start": "Website: \"Isolated VMs that start in under a second\" and \"Your custom image, running in under a second\" (vendor claims)",
      "cold_start_short": "<1 s (claim)",
      "max_session": "Maximum duration is set per session (default not stated); non-sticky sessions pause automatically at their maximum; sticky sessions run until terminated",
      "max_session_short": "Configurable; sticky = unlimited",
      "persistence": "Pause preserves memory and durable files (/home/tenki and volumes; /tmp survives a pause but is cleared on snapshot) and stops compute charges; snapshots capture disk and memory and can be forked; persistent volumes keep data across sessions",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Inbound and outbound networking on by default and fixed at create time; outbound allowlists up to 256 domains and 64 IPv4 CIDRs; `expose_port` creates authenticated, tunnelled preview URLs with optional TTL; no public IPs",
      "pricing": "Per second: vCPU $0.000014 ($0.0504/hour), memory $0.0000045 per GiB ($0.0162/GiB-hour), disk $0.00000003 per GiB ($0.000108/GiB-hour, first 5 GiB free), kept storage $0.000000076 per GiB-second. Starter $0; Team $250/month ($200/month billed yearly) with $100 monthly credits, up to 100 concurrent sessions and 16 cores/64 GB; Enterprise custom",
      "free_tier": "Starter: $100 one-time credit (granted when a payment method is added), up to 5 concurrent sessions, 4 cores/8 GB, 100 GiB storage",
      "compliance": "SOC 2 Type II through parent Luxor Technology's programme (report under NDA); compute in SOC 2-attested and ISO 27001-certified data centres; GDPR alignment with certification on the roadmap",
      "regions": "Multi-region availability mentioned; regions not listed (preview URL example uses a \"us\" subdomain)",
      "best_for": "Coding agents that need a full Linux VM with Docker and systemd, pausable between tasks, at per-second rates",
      "overview": [
        "Tenki Sandbox is the agent-VM product of Tenki Cloud, which describes itself as a product of its parent company Luxor Technology, built from Seattle, and which also sells CI runners. Each session is a fresh, single-tenant Linux VM with its own kernel (the hypervisor is not named) whose disk mounts directly from a Ceph cluster so nothing is copied before boot; the site claims sub-second starts. Defaults are 2 vCPU, 4,096 MB and a 5 GiB root disk, configurable at creation, with Docker, root and systemd available. SDKs exist for TypeScript, Python and Go alongside a CLI, and Claude Code and Codex are listed as supported harnesses.",
        "Sessions can be paused (memory and durable files kept, billing stops), snapshotted with memory and forked, and attached to persistent volumes; sticky sessions run until terminated while others pause at their maximum duration. Pricing is per second for vCPU, memory and disk, with a free Starter tier carrying a $100 credit and a Team plan with $100 of monthly credits. Tenki states that it operates under its parent company's SOC 2 Type II programme. Structured data on the page gives a software version of \"2026.11\"."
      ],
      "notable": [
        "Tenki Cloud founded 2024 in Seattle; parent company Luxor Technology; \"Built In Seattle — Featured Startup\".",
        "Pricing FAQ: \"Team is a $250/month subscription, or $200/month billed yearly\".",
        "Tenki for Startups offers $10K in free credits and up to $40K in matching grants across Tenki products."
      ],
      "pros": [
        "You want a full VM with Docker and systemd rather than a container-shaped sandbox.",
        "You need pause-and-resume with memory kept and no charge while paused.",
        "You want Go support alongside TypeScript and Python SDKs.",
        "You are already using Tenki runners and want one credit balance."
      ],
      "cons": [
        "Hypervisor, region list and default session duration are not documented.",
        "SOC 2 Type II is through parent company Luxor Technology's programme, with reports under NDA.",
        "Inbound and outbound networking are create-time settings and cannot be toggled afterwards."
      ],
      "verdict": "A per-second VM sandbox from a CI-runner vendor that runs its own hardware, suited to agents needing a full Linux VM; the hypervisor and region list are not documented.",
      "peers": [
        "e2b",
        "daytona",
        "superserve",
        "withruntime",
        "instavm",
        "namespace"
      ],
      "sources": [
        "https://tenki.cloud/products/sandbox",
        "https://tenki.cloud/docs/sandbox/quickstart",
        "https://tenki.cloud/docs/sandbox/sessions",
        "https://tenki.cloud/pricing",
        "https://tenki.cloud/company/security",
        "https://tenki.cloud/company/about.md",
        "https://tenki.cloud/docs/sandbox/concepts.md",
        "https://github.com/LuxorLabs"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "withruntime",
      "name": "Runtime",
      "company": "Runtime LLC",
      "hq": {
        "city": "Denver",
        "region": "Colorado",
        "country": "US",
        "source": "https://withruntime.com/about.md"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://withruntime.com",
      "docs": "https://withruntime.com/docs/llms.txt",
      "pricing_url": "https://withruntime.com/pricing.md",
      "github": "https://github.com/withruntime/runtime",
      "tagline": "Firecracker microVM sandboxes billed on active CPU at $0.025 per vCPU-hour, with pause that keeps memory, forks, SDKs in five languages and no plan fee.",
      "isolation": "Firecracker microVM with its own Linux kernel, disk and memory per sandbox, on dedicated servers Runtime operates",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Python",
        "Go",
        "Ruby",
        "Java",
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": "Runtime measures create and exec from outside every one to five minutes and publishes p50/p95/p99 on its speed page; the September 2026 figures were taken down on 27 October 2026, so no current figure is published",
      "cold_start_short": null,
      "max_session": "No maximum documented; a sandbox pauses itself after 300 seconds of inactivity by default (`idlePauseSeconds` 60–86,400, or 0 for never)",
      "max_session_short": "No limit documented",
      "persistence": "Pause keeps memory and processes; paused sandboxes are kept while the account has credit or for 1–365 days as set (30 days after credit runs out with a card or prior top-up, 7 days otherwise); snapshots of running or paused sandboxes; forks of 1–10 running copies with memory and processes; volumes backed up daily; S3, R2 and GCS bucket mounts",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Per-sandbox allow and deny lists, or internet off; sandboxes reach only the public internet, private addresses refused; HTTPS previews on runtimehost.com (private by default with rotatable tokens), custom domains, public TCP ports, WireGuard private networks and a dedicated outbound IPv4 ($5 per 30 days); secrets substituted by the host on outbound requests",
      "pricing": "$0.025 per active vCPU-hour and $0.0075 per reserved GiB-hour, billed per second; a waiting sandbox bills memory plus a 0.05 vCPU floor; paused sandboxes, snapshots and images $0.08 per GB per 30-day month; volumes about $0.11 per GiB per 30 days; outbound traffic free to 100 GiB a month then $0.02 per GB; no plan, seat or minimum fee",
      "free_tier": "100 free hours of a 1 GB machine (100 GiB-hours of memory, 25 vCPU-hours of active CPU) in the first 30 days with no card, renewing monthly once the account has paid; without credit up to 4 sandboxes at once of 2 vCPU and 4 GiB",
      "compliance": "No certification of Runtime itself is named; the trust page lists subprocessor attestations (Virginia data centre SOC 1/2/3 Type 2 and ISO 27001; Neon and Vercel SOC 2 Type 2); DPA included in every account; SAML/OIDC SSO and SCIM free; 99% monthly API uptime promise on paid accounts",
      "regions": "All processing in one region in Virginia, United States (trust page); limits are counted per region",
      "best_for": "Cost-sensitive agent workloads with many short runs, where active-CPU billing and a free monthly allowance matter",
      "overview": [
        "Runtime is a hosted sandbox service from Runtime LLC of Denver, Colorado. Each sandbox is a Firecracker microVM with its own kernel on servers Runtime operates in a Virginia data centre. The client side is unusually broad: SDKs for JavaScript, Python, Go, Ruby and Java, a CLI, an MCP server and an OpenAPI-described HTTP API, all Apache-2.0 on GitHub, plus a code interpreter for Python, JavaScript, TypeScript, R, Java, Bash and Go, desktop automation, Docker inside sandboxes, SSH, scheduled jobs, signed webhooks and OpenTelemetry export. Migration guides cover code written for E2B, Daytona, Vercel Sandbox, Blaxel, Morph, Modal and Cloudflare.",
        "The pricing model is the pitch: CPU is billed only while active at $0.025 per vCPU-hour, memory at $0.0075 per reserved GiB-hour, with no plan fee and a renewing free allowance of 100 GiB-hours; the vendor's own price index claims 42% to 88% lower cost than fourteen providers for a 1,000-run example job. Sandboxes pause after five minutes idle by default, keeping memory and processes, and can be forked into up to ten running copies. Governance features (SSO, SCIM, roles, 400-day audit log, DPA) are free, but the trust page names no certification of Runtime itself, listing only its subprocessors' attestations. The speed page currently shows no figures, having withdrawn them on 27 October 2026."
      ],
      "notable": [
        "Pricing page (checked 25 October 2026 by the vendor): example job of 1,000 one-minute runs costs $0.64 on Runtime against $1.11 to $5.33 elsewhere, a vendor comparison.",
        "Limits with credit: 100 sandboxes at once (200 with auto top-up), up to 16 vCPU, 64 GiB and 400 GiB disk each; 1,000 kept sandboxes.",
        "Client repository is Apache-2.0 with 2 GitHub stars at verification; payment accepted by card or USDC on Base, Solana or Tempo."
      ],
      "pros": [
        "You want the lowest published per-vCPU rate with billing only for active CPU.",
        "You need SDKs in Go, Ruby or Java as well as Python and TypeScript.",
        "You want SSO, SCIM, audit logging and a DPA without an enterprise contract.",
        "You run many short agent jobs and want a renewing free monthly allowance."
      ],
      "cons": [
        "Single region: everything runs in one region in Virginia, United States.",
        "No startup latency figures are currently published, and no certification of Runtime itself is named.",
        "Waiting sandboxes still bill memory plus a CPU floor, so idle time is not free unless paused."
      ],
      "verdict": "A Firecracker sandbox service billed on active CPU with no plan fee, running in a single Virginia region and naming no certification of its own; suited to cost-driven batch agent work.",
      "peers": [
        "e2b",
        "daytona",
        "superserve",
        "instavm",
        "leap0",
        "tenki-sandbox"
      ],
      "sources": [
        "https://withruntime.com",
        "https://withruntime.com/about.md",
        "https://withruntime.com/pricing.md",
        "https://withruntime.com/docs/limits.md",
        "https://withruntime.com/docs/speed.md",
        "https://withruntime.com/docs/trust",
        "https://withruntime.com/docs/llms.txt",
        "https://github.com/withruntime/runtime",
        "https://withruntime.com/docs/trust.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "temps",
      "name": "Temps",
      "company": "Temps (gotempsh)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://temps.sh",
      "docs": "https://temps.sh/docs/quickstart",
      "pricing_url": "https://temps.sh/pricing",
      "github": "https://github.com/gotempsh/temps",
      "tagline": "Self-hosted single-binary deployment platform (MIT or Apache-2.0) whose AI feature set includes Docker or Firecracker agent sandboxes with a browser terminal and preview URLs.",
      "isolation": "Docker container by default (capabilities dropped, no-new-privileges, shared host kernel); optional Firecracker microVM backend with its own guest kernel on Linux hosts with /dev/kvm, chosen per sandbox",
      "isolation_short": "Docker or Firecracker",
      "isolation_class": "mixed",
      "open_source": true,
      "license": "MIT OR Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "CLI",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Workspaces suspend when idle and wake on the next command; snapshots are not documented",
      "snapshots": null,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Password-protected preview URLs and an in-browser command runner; the platform proxy (built on Pingora) provides automatic HTTPS",
      "pricing": "Free to self-host, no usage billing; an optional Premium licence is sold separately (details on the pricing page)",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": "Your own server",
      "best_for": "Teams already self-hosting a PaaS who want built-in agent sandboxes next to their deployments",
      "overview": [
        "Temps is a self-hosted deployment platform shipped as a single Rust binary that bundles app deployments, analytics, error tracking, session replay, uptime monitoring, email, tracing, an AI gateway and AI agent sandboxes. It installs on a fresh Ubuntu server with one command and is dual-licensed MIT or Apache-2.0, with 915 GitHub stars at verification. Sandboxes are one feature among many: isolated environments driven from the CLI or REST API, with a live in-browser terminal and preview URLs; a Node sandbox SDK exists in the repository but is not yet on npm.",
        "The default backend is a Docker container with capabilities dropped and no-new-privileges set, sharing the host kernel. On Linux hosts with `/dev/kvm`, `temps firecracker setup` enables a Firecracker backend that boots each sandbox as a microVM with its own guest kernel; the backend is chosen per sandbox and a Firecracker request fails rather than silently falling back, and macOS hosts cannot enable it. Workspaces suspend when idle and wake on the next command. Snapshots, resource limits, GPU and session limits are not documented, and the vendor's own comparison page calls E2B \"the more mature product\"."
      ],
      "notable": [
        "Dual-licensed MIT or Apache-2.0; 915 GitHub stars and 4,353 commits at verification; \"55+ Rust crates\".",
        "Firecracker backend requires Linux x86_64 or aarch64 with /dev/kvm and a one-time setup; Docker backend works on macOS for local development.",
        "Anonymous usage telemetry is on by default and can be disabled with `TEMPS_TELEMETRY=0`."
      ],
      "pros": [
        "You already run or want a self-hosted PaaS and would like sandboxes without another service.",
        "You want to choose Docker or Firecracker isolation per sandbox on one host.",
        "You need agent sandboxes with no per-second or per-sandbox charges."
      ],
      "cons": [
        "Sandboxes are a feature of a deployment platform, not a dedicated sandbox product; the vendor calls E2B more mature.",
        "No SDK on npm yet; interaction is via CLI and REST API.",
        "Snapshots, resource limits, session limits and GPU are not documented.",
        "Default Docker backend shares the host kernel; Firecracker needs a KVM-capable Linux host."
      ],
      "verdict": "A self-hosted deployment platform with built-in Docker or Firecracker agent sandboxes, suited to teams that want sandboxes alongside the rest of the Temps platform.",
      "peers": [
        "docker-sandboxes",
        "microsandbox",
        "pullrun",
        "coder",
        "opensandbox"
      ],
      "sources": [
        "https://temps.sh",
        "https://temps.sh/agents",
        "https://github.com/gotempsh/temps",
        "https://temps.sh/pricing",
        "https://raw.githubusercontent.com/gotempsh/temps/main/README.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "archil",
      "name": "Archil",
      "company": "Archil, Inc.",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://archil.com/privacy"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://archil.com",
      "docs": "https://docs.archil.com",
      "pricing_url": "https://archil.com/pricing",
      "github": null,
      "tagline": "Firecracker microVM sandboxes and one-shot serverless commands that run against durable, S3-backed Archil disks, with forks and pause/resume.",
      "isolation": "Firecracker microVM with its own Linux kernel for each persistent sandbox; each serverless command also runs in its own fresh Linux microVM colocated with the disk",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": true,
      "sdks": [
        "JavaScript/TypeScript",
        "Python",
        "CLI",
        "REST API"
      ],
      "cold_start": "Serverless commands: median boot-plus-mount (queueMs) of 37 ms and 82 ms end to end for a trivial command, measured by Archil from a laptop to a disk in AWS us-west-2; the first call on a new disk took a median 331 ms. Boot time for persistent sandboxes is not documented",
      "cold_start_short": "37 ms exec (claim)",
      "max_session": "Persistent sandboxes: maxTtlSeconds defaults to and is capped at 24 hours, after which the sandbox is paused, not deleted; Free plan sessions last at most 30 minutes. Serverless commands stop after 5 minutes",
      "max_session_short": "24 h (then paused)",
      "persistence": "Each sandbox's root filesystem is ext4 on an Archil disk, so files and installed packages survive stop, start, pause and resume; pause also keeps memory and running processes. Mounted Archil disks outlive the sandbox, are SSD-cached over object storage (your own S3 bucket or Archil-managed) and are readable over POSIX and the S3 API",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Free plan egress is limited to a managed allowlist of package registries, Git hosts, container registries, OS mirrors and model provider APIs; on paid plans you set a per-sandbox firewall policy, and a sandbox created without one has unrestricted egress. Ports are private until exposed; exposed ports are public TLS (SNI) endpoints without authentication, or can be gated with port tokens sent in an X-Archil-Token header",
      "pricing": "Free: $0, 5 file systems, 10 GB per disk, 5 concurrent sandboxes, 1,000 serverless tool calls, 1 vCPU-hour and 2 GiB-hours of sandbox compute a month (hard caps). Developer: $10 monthly minimum, 20 concurrent sandboxes. Team: $500 monthly minimum, 250 concurrent sandboxes. Usage on paid plans: $0.055 per vCPU-hour and $0.020 per GiB-hour while a sandbox runs, $0.05 per 1,000 serverless calls, $0.20 per GiB-month active storage, $0.07 per GB egress. Enterprise: custom",
      "free_tier": "Free plan: 1 vCPU-hour and 2 GiB-hours of sandbox compute, 1,000 serverless tool calls and 5 disks of up to 10 GB a month, no card required",
      "compliance": "SOC 2 Type II and HIPAA listed on the website; encryption at rest and in transit",
      "regions": "AWS us-east-1, us-west-2 and eu-west-1; BYOC deployments on AWS, Google Cloud, Azure or bare metal",
      "best_for": "Agents whose files must outlive the sandbox: per-user or per-session disks that serverless commands and long-running microVMs share",
      "overview": [
        "Archil describes itself as \"Sandboxes and durable disks for AI\". The product is built around the Archil disk: a file system that keeps hot data on SSD and the rest in object storage, either your own S3 bucket (where files stay ordinary S3 objects) or storage Archil manages, and that can be read over POSIX or the S3 API or mounted on your own hosts with the Archil CLI. Code runs against those disks in two ways. Serverless execution runs one Bash, Python or Node.js command in a fresh microVM next to the data and returns, with a 5-minute limit per command. Persistent sandboxes are Firecracker microVMs with their own kernel and a persistent root disk that run for up to 24 hours per session.",
        "A persistent sandbox takes 1 to 32 vCPU and 0.25 to 64 GiB of memory on x86-64, boots from a public or private OCI image, and can be stopped, started, paused (keeping memory and processes) or forked; Archil says a fork's checkpoint takes about 15 ms and its branch about 35 ms regardless of disk size. Idle sandboxes pause after 30 minutes by default, and traffic to an exposed port wakes them. There are TypeScript and Python SDKs, a CLI, a REST API and an AI SDK integration. Compute is billed per vCPU-hour and GiB-hour only while running, with storage billed separately. GPU support is not documented, and the product is not open source, although BYOC deployments are offered."
      ],
      "notable": [
        "Docs: forking a sandbox checkpoints its root disk in about 15 ms and creates a branch in about 35 ms, regardless of disk size.",
        "Serverless execution: median 37 ms boot plus mount per command (vendor measurement); hundreds of commands can run in parallel.",
        "Disks can sit on your own S3 bucket, where files remain ordinary S3 objects readable by other tools."
      ],
      "pros": [
        "You want an agent's files on durable disks that outlive any one sandbox and are readable over S3.",
        "You need both one-shot tool calls and long-running microVMs against the same data.",
        "You want to fork a sandbox's disk in milliseconds to branch an agent's work.",
        "You need sandboxes that pause when idle and wake on incoming traffic, billed only while running."
      ],
      "cons": [
        "A persistent sandbox session is capped at 24 hours before it is paused.",
        "Boot time for persistent sandboxes and GPU support are not documented.",
        "x86-64 only, and processes run as root inside the VM.",
        "Not open source; self-managed deployment is offered as BYOC rather than a public release."
      ],
      "verdict": "A storage-first sandbox: Firecracker microVMs and one-shot commands built around durable, forkable disks.",
      "peers": [
        "fly-sprites",
        "e2b",
        "morph",
        "freestyle",
        "opencomputer",
        "daytona"
      ],
      "sources": [
        "https://archil.com",
        "https://docs.archil.com",
        "https://docs.archil.com/compute/sandboxes/introduction",
        "https://docs.archil.com/compute/serverless-execution",
        "https://archil.com/pricing",
        "https://docs.archil.com/administration/billing",
        "https://archil.com/privacy"
      ],
      "verified_on": "2026-11-16"
    },
    {
      "id": "isorun",
      "name": "Isorun",
      "company": "Isorun, Inc.",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://isorun.ai",
      "docs": "https://docs.isorun.ai",
      "pricing_url": "https://docs.isorun.ai/getting-started/pricing",
      "github": "https://github.com/isorun/sdk",
      "tagline": "Hosted KVM microVM sandboxes booted from pinned OCI images, with fork of a running VM, hibernate/resume, host-side egress rules and credential injection.",
      "isolation": "KVM microVM per sandbox with its own guest kernel; the docs do not name the hypervisor. Egress policy and the credential proxy are enforced on the host, outside the guest",
      "isolation_short": "KVM microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "REST API",
        "MCP server"
      ],
      "cold_start": "Homepage: VMs \"ready in 10ms\" (vendor claim); the API reference example shows create_ms of 9. The first time a runner sees an image it pulls and prepares it, which the docs say takes 10 to 60 seconds",
      "cold_start_short": "10 ms (claim)",
      "max_session": "No maximum lifetime documented. Each sandbox has an idle auto-destroy timer, 300 seconds by default, set per create with timeoutSec or extended with setTimeout(); passing 0 disables it",
      "max_session_short": "300 s idle timeout (default)",
      "persistence": "The image is mounted read-only with a writable scratch layer (4 GiB by default) that lasts for the sandbox's lifetime, including across hibernate/resume. Hibernate keeps filesystem, memory and running processes. Snapshots persist until deleted and restore on any runner in the same region. Optional named persistent disks (ext4 on Tigris Cloud Disk with a local NVMe cache) mount at /workspace, outlive sandboxes and can be snapshotted and forked copy-on-write",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Egress is unrestricted by default; per-sandbox allow/deny rules (CIDRs, IPs, hostnames matched on SNI, wildcards) or named profiles (unrestricted, locked-down, claude-code, openai) are enforced at the host boundary. A host-side proxy injects API keys for OpenAI, Anthropic, Google, GitHub and Hugging Face so the guest only sees placeholders, with optional method and path rules. sandbox.url(port) proxies any port over HTTPS and WebSocket; the docs differ on whether these URLs require the API key",
      "pricing": "Per-second billing with no minimums: $0.025 per vCPU-hour plus $0.015 per GiB-hour of RAM ($0.040 an hour for the default 1 vCPU + 1 GiB; 8 vCPU + 16 GiB is $0.440 an hour). Hibernated sandboxes are not billed for runtime. Persistent disks: $0.02 per GB-month of data stored. The homepage says egress traffic and platform fees are included",
      "free_tier": "$50 of credit on sign-up, no card required (about 1,250 hours of a 1 vCPU + 1 GiB sandbox)",
      "compliance": "SOC 2 Type II listed as \"(soon)\" on the homepage; each sandbox has an HMAC-chained, tamper-evident audit log",
      "regions": "Organizations are scoped to a US or EU region at sign-up; the API key routes to that region's runner",
      "best_for": "Agents that need full Linux VMs from any pinned OCI image, with fork, hibernate and API keys kept out of the guest",
      "overview": [
        "Isorun is a hosted sandbox service in which each sandbox is a KVM microVM with its own guest kernel, booted from an OCI image from Docker Hub, GHCR or a private registry. Floating tags such as latest, lts and stable are rejected, so images must be pinned to a tag or digest. A sandbox defaults to 1 vCPU, 1 GiB RAM and a 4 GiB scratch disk, and is destroyed after 300 idle seconds unless the timer is changed. The homepage says sandboxes are \"ready in 10ms\"; the first boot of a new image on a runner takes 10 to 60 seconds while it is prepared. The docs do not name the hypervisor.",
        "State features include fork, which clones a running sandbox's filesystem, memory and processes into N children, snapshot and restore within a region, hibernate/resume with processes kept at the same PIDs, and optional persistent disks on Tigris that can be snapshotted and forked. Security features run on the host: per-sandbox egress allow/deny rules and named profiles, a credential proxy that injects API keys so the guest sees only placeholders, and an HMAC-chained audit log. There is a TypeScript SDK (MIT, Node 20+), a REST API, an MCP server, adapters for OpenAI Agents and LangChain.js, and an adapter that runs Claude Managed Agents sessions in Isorun microVMs; a Python SDK is described as \"in development\". The homepage calls the service a public beta, while the changelog lists v1.0.0 as the public launch on 16 June 2026."
      ],
      "notable": [
        "Fork clones a running sandbox, including memory and process state, into N children with a single call.",
        "API keys for OpenAI, Anthropic, Google, GitHub and Hugging Face are injected by a host-side proxy; the guest's environment holds only placeholders.",
        "Ships a self-hosted sandbox adapter for Claude Managed Agents that binds each Anthropic session to its own microVM."
      ],
      "pros": [
        "You want each sandbox to be a microVM with its own kernel, booted from any pinned OCI image.",
        "You need to fork a running sandbox with its memory intact to branch an agent's work.",
        "You want third-party API keys kept outside the VM and egress restricted at the host.",
        "You want per-second billing with no charge while a workspace is hibernated."
      ],
      "cons": [
        "Only a TypeScript SDK today; the Python SDK is described as in development.",
        "The hypervisor, GPU support and a maximum session length are not documented.",
        "Egress is unrestricted unless you set a policy or profile at create time."
      ],
      "verdict": "A KVM microVM sandbox service centred on forking, hibernation and host-side secret handling, with a TypeScript-first SDK.",
      "peers": [
        "e2b",
        "morph",
        "fly-sprites",
        "archil",
        "daytona",
        "freestyle"
      ],
      "sources": [
        "https://isorun.ai",
        "https://docs.isorun.ai",
        "https://docs.isorun.ai/getting-started/quickstart",
        "https://docs.isorun.ai/getting-started/authentication",
        "https://docs.isorun.ai/getting-started/pricing",
        "https://docs.isorun.ai/sandboxes/lifecycle",
        "https://docs.isorun.ai/sandboxes/persistent-storage",
        "https://docs.isorun.ai/sandboxes/checkpoints-and-rollback",
        "https://docs.isorun.ai/sandboxes/public-urls",
        "https://docs.isorun.ai/sandboxes/any-oci-image",
        "https://docs.isorun.ai/security/network-filtering",
        "https://docs.isorun.ai/security/credential-injection",
        "https://docs.isorun.ai/sdks/typescript",
        "https://docs.isorun.ai/integrations/crewai",
        "https://docs.isorun.ai/api-reference",
        "https://docs.isorun.ai/changelog",
        "https://github.com/isorun/sdk"
      ],
      "verified_on": "2026-11-18"
    },
    {
      "id": "aws-agentcore",
      "name": "AWS Bedrock AgentCore",
      "company": "Amazon Web Services",
      "hq": {
        "city": "Seattle",
        "region": "Washington",
        "country": "US",
        "source": "https://aws.amazon.com/privacy/"
      },
      "kind": "hyperscaler",
      "tier": 2,
      "website": "https://aws.amazon.com/bedrock/agentcore/",
      "docs": "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/code-interpreter-tool.html",
      "pricing_url": "https://aws.amazon.com/bedrock/agentcore/pricing/",
      "github": null,
      "tagline": "AWS’s managed agent platform: a Code Interpreter tool for running agent-generated code and a Runtime that hosts each agent session in its own microVM.",
      "isolation": "Runtime: a dedicated microVM per user session with isolated CPU, memory and filesystem, sanitized after the session. Code Interpreter: documented as a containerized environment within AgentCore",
      "isolation_short": "MicroVM (Runtime)",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": null,
      "sdks": [
        "Python (AgentCore SDK)",
        "AWS SDKs",
        "AWS CLI"
      ],
      "cold_start": "No numeric claim; Runtime v2 restores from snapshots for consistent cold starts",
      "cold_start_short": null,
      "max_session": "Code Interpreter: 15 min default, up to 8 hours. Runtime microVMs: up to 8 hours, stop after 15 min idle",
      "max_session_short": "Up to 8 h",
      "persistence": "State persists between executions within a session; Runtime can persist filesystem across stop/resume (managed session storage in preview). Large files via S3",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented for Code Interpreter or Runtime microVMs; the separate Runtime “Instances” compute type is described for agents that need GPUs",
      "networking": "Code Interpreter network modes: Sandbox (isolated) or Public; VPC connectivity supported",
      "pricing": "Code Interpreter: $0.0895 per vCPU-hour + $0.00945 per GB-hour, billed per second; CPU billed on actual consumption (not during I/O wait), memory on peak",
      "free_tier": "No product-specific free tier is listed; general AWS new-customer credits apply",
      "compliance": "HIPAA eligible and pursuing FedRAMP per the AWS FAQ; AWS reports an internal assessment against programmes including ISO, PCI and SOC, with third-party audits in upcoming cycles",
      "regions": "15 AWS Regions across North America, South America, Europe and Asia Pacific",
      "best_for": "Teams already on AWS that need IAM- and VPC-integrated code execution for agents",
      "overview": [
        "Amazon Bedrock AgentCore is AWS’s set of managed building blocks for production agents. Two pieces matter for sandboxing: the **Code Interpreter** tool, which runs Python, JavaScript, TypeScript and shell commands for an agent, and the **Runtime**, which hosts the agent itself with each user session in a dedicated microVM.",
        "Its distinguishing features are long sessions (up to 8 hours), consumption-based CPU billing that does not charge for I/O wait such as waiting on model responses, and native IAM, VPC and regional controls. It is model- and framework-agnostic and supports MCP and A2A."
      ],
      "notable": [
        "Code Interpreter sessions can run up to 8 hours — long for a hosted code interpreter.",
        "CPU is not billed while the agent waits on I/O, such as LLM responses."
      ],
      "pros": [
        "You are already on AWS and want IAM, VPC and CloudTrail integration",
        "You need long-running sessions (up to 8 hours)",
        "You want consumption-based CPU billing that skips I/O wait"
      ],
      "cons": [
        "AWS-only; no self-hosting or other clouds",
        "Code Interpreter isolation is documented only as “containerized”",
        "GPU support is documented only for the Runtime “Instances” compute type, not for Code Interpreter or microVMs"
      ],
      "verdict": "A managed AWS service for hosting agents and running their code, suited to teams already on AWS that want IAM and VPC integration; it runs only on AWS.",
      "peers": [
        "azure-dynamic-sessions",
        "google-code-execution",
        "e2b",
        "modal"
      ],
      "extra_vs": [
        "e2b"
      ],
      "sources": [
        "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/code-interpreter-tool.html",
        "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/code-interpreter-resource-session-management.html",
        "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-how-it-works.html",
        "https://aws.amazon.com/bedrock/agentcore/pricing/",
        "https://aws.amazon.com/bedrock/agentcore/faqs/",
        "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-persistent-filesystems.html"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "azure-dynamic-sessions",
      "name": "Azure Container Apps dynamic sessions",
      "short_name": "Azure dynamic sessions",
      "company": "Microsoft",
      "hq": {
        "city": "Redmond",
        "region": "Washington",
        "country": "US",
        "source": "https://www.microsoft.com/en-us/about/office-locations"
      },
      "kind": "hyperscaler",
      "tier": 2,
      "website": "https://learn.microsoft.com/en-us/azure/container-apps/sessions",
      "docs": "https://learn.microsoft.com/en-us/azure/container-apps/sessions",
      "pricing_url": "https://azure.microsoft.com/en-us/pricing/details/container-apps/",
      "github": null,
      "tagline": "Azure service that hands out Hyper-V-isolated sessions from prewarmed pools, either as a built-in code interpreter or running your own container image.",
      "isolation": "Hyper-V isolation; each session is isolated by a Hyper-V boundary",
      "isolation_short": "Hyper-V VM",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "REST API",
        "Azure CLI",
        "LangChain",
        "LlamaIndex",
        "Semantic Kernel"
      ],
      "cold_start": "Prewarmed pools allocate sessions in milliseconds (no specific figure)",
      "cold_start_short": "Sub-second (pooled)",
      "max_session": "Code interpreter: 220 s per execution; sessions deleted after a 5–60 min idle cooldown",
      "max_session_short": "220 s / exec",
      "persistence": "Session reuse by identifier; files in /mnt/data for the session’s life. Snapshots are not documented; sessions are ephemeral",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Per-pool egress setting; egress disabled by default",
      "pricing": "Code interpreter: $0.03 per session-hour (East US, Azure Retail Prices API), billed in one-hour increments. Custom container sessions billed on the Dedicated plan",
      "free_tier": "None stated for dynamic sessions",
      "compliance": null,
      "regions": "38 Azure regions listed in the docs",
      "best_for": "Azure-native apps running LLM-generated Python or Node code with egress off by default",
      "overview": [
        "Azure Container Apps dynamic sessions provide sandboxed execution from prewarmed pools. The built-in code interpreter runs Python, Node.js and shell; custom container sessions run any image. Each session sits behind a Hyper-V boundary and outbound network access is disabled by default.",
        "First-party integrations with LangChain, LlamaIndex and Semantic Kernel make it straightforward to add a code tool to Azure-hosted agents."
      ],
      "notable": [
        "Up to 600 concurrent sessions per code interpreter pool.",
        "Code interpreter sessions are billed per allocated session in increments of one hour."
      ],
      "pros": [
        "You build on Azure and want Hyper-V isolation without managing infrastructure",
        "You want egress disabled by default",
        "You use LangChain, LlamaIndex or Semantic Kernel"
      ],
      "cons": [
        "220-second execution limit on the code interpreter",
        "Code interpreter sessions are billed in one-hour increments",
        "Snapshots, pause/resume and GPU support are not documented for sessions"
      ],
      "verdict": "A Hyper-V-isolated code interpreter and custom-container session service for teams on Azure, with outbound network access off by default; Microsoft describes sessions as ephemeral and designed for short-lived tasks.",
      "peers": [
        "aws-agentcore",
        "google-code-execution",
        "openai-code-interpreter",
        "e2b"
      ],
      "sources": [
        "https://learn.microsoft.com/en-us/azure/container-apps/sessions",
        "https://learn.microsoft.com/en-us/azure/container-apps/sessions-code-interpreter",
        "https://learn.microsoft.com/en-us/azure/container-apps/session-pool",
        "https://azure.microsoft.com/en-us/pricing/details/container-apps/",
        "https://learn.microsoft.com/en-us/azure/container-apps/billing",
        "https://learn.microsoft.com/en-us/azure/container-apps/sessions-usage"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "google-code-execution",
      "name": "Google Agent Platform Code Execution",
      "short_name": "Agent Platform Code Execution",
      "company": "Google Cloud",
      "hq": {
        "city": "Mountain View",
        "region": "California",
        "country": "US",
        "source": "https://policies.google.com/terms"
      },
      "kind": "hyperscaler",
      "tier": 3,
      "website": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/sandbox",
      "docs": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/sandbox/code-execution-overview",
      "pricing_url": "https://cloud.google.com/products/gemini-enterprise-agent-platform/pricing",
      "github": null,
      "tagline": "Google-managed sandbox API (the docs also refer to it under Vertex AI Agent Engine) that lets agents deployed anywhere run Python or JavaScript in stateful, network-isolated environments.",
      "isolation": "gVisor-isolated container in a Google-managed tenant project; does not inherit the calling agent’s credentials",
      "isolation_short": "gVisor",
      "isolation_class": "gvisor",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "Python (Agent Platform SDK)",
        "REST API"
      ],
      "cold_start": "Docs: create and execute code “in under a second”; a cold provision can take up to 2 minutes",
      "cold_start_short": "<1 s (warm)",
      "max_session": "State retained up to 14 days (configurable TTL)",
      "max_session_short": "14-day state",
      "persistence": "Stateful across calls for the TTL; snapshots not supported for Code Execution sandboxes",
      "snapshots": false,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "No network access from the code execution sandbox; VPC Service Controls supported",
      "pricing": "$0.085 per vCPU-hour + $0.009 per GiB-hour, rounded to the second",
      "free_tier": null,
      "compliance": "VPC Service Controls, CMEK, data residency at rest and HIPAA per Google’s Agent Platform security table",
      "regions": null,
      "best_for": "Google Cloud agents that need stateful, no-network Python execution",
      "overview": [
        "Google’s managed Code Execution sandbox, part of the Gemini Enterprise Agent Platform (the docs also refer to Vertex AI Agent Engine), runs Python or JavaScript in gVisor-isolated containers with no network access. State persists across calls for up to 14 days, which suits conversational analysis tools.",
        "The library set is fixed — you cannot install your own packages."
      ],
      "notable": [
        "Agents do not have to run on Google Cloud to call it.",
        "File I/O up to 100 MB per request or response."
      ],
      "pros": [
        "You want a no-network, gVisor-isolated Python tool",
        "You need state that lasts days, not minutes",
        "You are on Google Cloud and need CMEK or VPC Service Controls"
      ],
      "cons": [
        "Python and JavaScript only, with a fixed library set",
        "No network access and no snapshots"
      ],
      "verdict": "A managed code-execution sandbox for Python and JavaScript with no network access, a fixed library set and state kept for up to 14 days, suited to data-analysis style agent tools.",
      "peers": [
        "gke-agent-sandbox",
        "aws-agentcore",
        "azure-dynamic-sessions",
        "openai-code-interpreter"
      ],
      "sources": [
        "https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/sandbox/code-execution-overview",
        "https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/sandbox",
        "https://cloud.google.com/products/gemini-enterprise-agent-platform/pricing",
        "https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale",
        "https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/sandbox/code-execution-quickstart"
      ],
      "verified_on": "2026-11-15"
    },
    {
      "id": "openai-code-interpreter",
      "name": "OpenAI Code Interpreter & Hosted Shell",
      "short_name": "OpenAI Code Interpreter",
      "company": "OpenAI",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://openai.com/policies/terms-of-use/"
      },
      "kind": "hyperscaler",
      "tier": 3,
      "website": "https://developers.openai.com/api/docs/guides/tools-code-interpreter",
      "docs": "https://developers.openai.com/api/docs/guides/tools-code-interpreter",
      "pricing_url": "https://developers.openai.com/api/docs/pricing",
      "github": null,
      "tagline": "Hosted containers in the OpenAI Responses API that the model uses to run Python (Code Interpreter) or shell commands (Hosted Shell).",
      "isolation": "Documented as a fully sandboxed virtual machine; underlying technology not named",
      "isolation_short": "VM (unspecified)",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "OpenAI Responses API",
        "Containers API",
        "OpenAI SDKs"
      ],
      "cold_start": null,
      "max_session": "Containers expire after 20 minutes of inactivity and cannot be reactivated",
      "max_session_short": "20 min idle",
      "persistence": "Files and state persist while the container is active; reusable by container ID; snapshots are not documented",
      "snapshots": null,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": null,
      "networking": "No outbound network by default; Hosted Shell can allow specific domains via an org allowlist and per-request network policy, with credential placeholders",
      "pricing": "Per container per 20-minute session: $0.03 (1 GB), $0.12 (4 GB), $0.48 (16 GB), $1.92 (64 GB); model tokens billed separately",
      "free_tier": null,
      "compliance": null,
      "regions": null,
      "best_for": "Apps already on the OpenAI Responses API that want model-driven code execution with no extra vendor",
      "overview": [
        "OpenAI’s hosted containers back two Responses API tools: **Code Interpreter** (Python) and **Hosted Shell** (a Debian environment with Python, Node.js, Java, Go and more). The model decides when to run code; you manage containers through the Containers API.",
        "It needs no additional vendor if you already use OpenAI models; it is available only through OpenAI's API, and containers expire after 20 minutes of inactivity."
      ],
      "notable": [
        "Memory tiers from 1 GB to 64 GB, fixed for the container’s life.",
        "Containers API limited to 100 requests per minute per organization."
      ],
      "pros": [
        "You already use the OpenAI Responses API",
        "You want zero additional infrastructure",
        "You need egress off by default with domain allowlisting"
      ],
      "cons": [
        "Available only through OpenAI’s API and models",
        "Containers expire after 20 idle minutes",
        "Isolation technology not documented"
      ],
      "verdict": "Hosted containers for apps on the OpenAI Responses API that want model-driven code execution without another vendor; containers expire after 20 minutes of inactivity.",
      "peers": [
        "e2b",
        "azure-dynamic-sessions",
        "google-code-execution",
        "riza"
      ],
      "sources": [
        "https://developers.openai.com/api/docs/guides/tools-code-interpreter",
        "https://developers.openai.com/api/docs/guides/tools-shell",
        "https://developers.openai.com/api/docs/pricing"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "gke-agent-sandbox",
      "name": "Kubernetes Agent Sandbox",
      "company": "Kubernetes SIG Apps · Google Cloud (GKE)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "website": "https://agent-sandbox.sigs.k8s.io",
      "docs": "https://docs.cloud.google.com/kubernetes-engine/docs/concepts/machine-learning/agent-sandbox",
      "pricing_url": null,
      "github": "https://github.com/kubernetes-sigs/agent-sandbox",
      "tagline": "Open-source Kubernetes controller and CRDs for isolated, stateful sandbox pods for AI agents, with first-class support on GKE.",
      "isolation": "Delegates to the cluster RuntimeClass — typically gVisor or Kata Containers. On GKE: GKE Sandbox (gVisor) or microVMs. Without a sandbox runtime, pods have only container isolation",
      "isolation_short": "gVisor or Kata",
      "isolation_class": "gvisor",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": true,
      "sdks": [
        "Kubernetes CRDs",
        "Python",
        "kubectl"
      ],
      "cold_start": "GKE docs: under one second with a SandboxWarmPool",
      "cold_start_short": "<1 s (warm pool)",
      "max_session": null,
      "persistence": "Persistent storage across restarts; pause/resume lifecycle. On GKE, Pod snapshots (which Google says might be in Preview)",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "On GKE, GPUs and TPUs with the gVisor runtime on supported models",
      "networking": "Default-deny NetworkPolicy created per SandboxTemplate",
      "pricing": "Free and open source; you pay for the underlying cluster compute",
      "free_tier": "Open source",
      "compliance": null,
      "regions": "Any Kubernetes cluster; any GKE region",
      "best_for": "Platform teams that want to run agent sandboxes on their own Kubernetes or GKE clusters",
      "overview": [
        "Agent Sandbox is a Kubernetes SIG Apps project that adds `Sandbox`, `SandboxTemplate`, `SandboxClaim` and `SandboxWarmPool` resources for running isolated, singleton, stateful pods for agents. Warm pools hand out pre-started sandboxes on demand.",
        "Isolation comes from whatever sandbox runtime your cluster uses; on GKE that is gVisor (GKE Sandbox) or microVMs, with Pod snapshots for suspend and resume."
      ],
      "notable": [
        "v1.0 released; latest v1.0.11 published 2026-11-12.",
        "Default-deny network policy for templated sandboxes."
      ],
      "pros": [
        "You already run Kubernetes and want sandboxes inside your own cluster",
        "You need GPUs or TPUs alongside sandboxing (on GKE with gVisor)",
        "You want an open, vendor-neutral API"
      ],
      "cons": [
        "You operate the controller, runtime and cluster yourself",
        "Isolation depends on the runtime you configure; without a sandbox runtime, Pods run without kernel isolation",
        "Requires a Kubernetes cluster (1.25 or later) with the controller and CRDs installed"
      ],
      "verdict": "An open-source, Kubernetes-native API and controller for self-hosting agent sandboxes, suited to platform teams that already run Kubernetes or GKE.",
      "peers": [
        "microsandbox",
        "daytona",
        "e2b",
        "northflank"
      ],
      "sources": [
        "https://github.com/kubernetes-sigs/agent-sandbox",
        "https://agent-sandbox.sigs.k8s.io",
        "https://docs.cloud.google.com/kubernetes-engine/docs/concepts/machine-learning/agent-sandbox",
        "https://github.com/kubernetes-sigs/agent-sandbox/releases"
      ],
      "verified_on": "2026-11-15"
    },
    {
      "id": "microsandbox",
      "extra_vs": [
        "e2b"
      ],
      "name": "Microsandbox",
      "company": "Super Rad Company",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "website": "https://microsandbox.dev",
      "docs": "https://docs.microsandbox.dev",
      "pricing_url": "https://microsandbox.dev/pricing",
      "github": "https://github.com/superradcompany/microsandbox",
      "tagline": "Open-source, local-first microVM runtime that gives each sandbox its own Linux kernel, with an optional managed cloud.",
      "isolation": "Hardware-virtualized microVMs built on libkrun; requires KVM on Linux, Apple Silicon on macOS or Windows Hypervisor Platform",
      "isolation_short": "MicroVM (libkrun)",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": true,
      "sdks": [
        "TypeScript",
        "Python",
        "Rust",
        "Go",
        "Ruby",
        "CLI",
        "MCP server"
      ],
      "cold_start": "README: average boot under 100 ms (measured on an M1 machine)",
      "cold_start_short": "<100 ms",
      "max_session": null,
      "persistence": "Volumes, snapshots and forking; detached sandboxes",
      "snapshots": true,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Network policy with hostname allowlists, port rules and private-network/metadata blocking; secrets substituted via placeholders so keys never enter the VM",
      "pricing": "Free to self-host. Managed cloud (private beta, access by request): free trial with a $10 one-time credit; Builder $20/month and Teams $200/month, each including the same amount of usage credit; usage billed at $0.048 per vCPU-hour and $0.003 per GiB-hour of memory; Enterprise custom",
      "free_tier": "Self-hosting is free; cloud free trial with a $10 one-time credit (1 sandbox at a time, 1-hour sessions)",
      "compliance": null,
      "regions": null,
      "best_for": "Self-hosted microVM isolation on developer machines or your own servers",
      "overview": [
        "Microsandbox runs OCI images inside lightweight microVMs on Linux, macOS (Apple Silicon) and Windows, giving each sandbox its own kernel without a cloud dependency. It ships SDKs in five languages, a CLI and an MCP server.",
        "Its network policy and placeholder-based secret injection are notably thoughtful for agent workloads. A managed cloud is in private beta; the project itself is labelled beta."
      ],
      "notable": [
        "Labelled beta; latest release v0.7.15.",
        "States Y Combinator backing."
      ],
      "pros": [
        "You want microVM isolation on your own laptop or servers",
        "You want secrets kept out of the VM via placeholder substitution",
        "You want SDKs beyond Python and TypeScript"
      ],
      "cons": [
        "Beta software",
        "GPU support is not documented",
        "Managed cloud is in private beta, with access by request"
      ],
      "verdict": "An open-source, beta-labelled microVM runtime for running sandboxes on your own machine or servers, suited to teams that want local or self-hosted isolation.",
      "peers": [
        "anthropic-srt",
        "docker-sandboxes",
        "gke-agent-sandbox",
        "e2b"
      ],
      "sources": [
        "https://github.com/superradcompany/microsandbox",
        "https://docs.microsandbox.dev",
        "https://microsandbox.dev",
        "https://microsandbox.dev/pricing",
        "https://github.com/superradcompany/microsandbox/releases"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "docker-sandboxes",
      "extra_vs": [
        "anthropic-srt"
      ],
      "name": "Docker Sandboxes",
      "company": "Docker, Inc.",
      "hq": {
        "city": "Palo Alto",
        "region": "California",
        "country": "US",
        "source": "https://www.docker.com/company/contact/"
      },
      "kind": "local",
      "tier": 2,
      "website": "https://docs.docker.com/ai/sandboxes/",
      "docs": "https://docs.docker.com/ai/sandboxes/",
      "pricing_url": "https://docs.docker.com/agentic-platform/signup/#billing",
      "github": "https://github.com/docker/sbx-releases",
      "tagline": "Docker’s `sbx` CLI for running AI coding agents in microVM sandboxes on your machine, with an experimental Docker-hosted cloud.",
      "isolation": "MicroVM with a separate kernel per sandbox, each running its own private Docker Engine",
      "isolation_short": "MicroVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": true,
      "byoc": false,
      "sdks": [
        "sbx CLI",
        "TypeScript (cloud)",
        "REST API (cloud)"
      ],
      "cold_start": null,
      "max_session": "Local: persist until removed. Cloud: governed by configured time-to-live",
      "max_session_short": "Until removed (local)",
      "persistence": "Local sandboxes keep packages and agent state until removed; cloud supports snapshots and volumes; `sbx move` copies a sandbox between local and cloud",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Local only, GPU passthrough on supported hosts; not in cloud sandboxes",
      "networking": "Deny-by-default network policy; outbound TCP proxied through the host; credentials injected by a host-side proxy so raw keys never enter the VM",
      "pricing": "CLI and local sandboxes free, including commercial use. Cloud sandboxes metered per second; per-unit rates are not published in the public docs",
      "free_tier": "Local use free; one-time credit for cloud sign-ups",
      "compliance": null,
      "regions": null,
      "best_for": "Running coding agents unsupervised on a laptop with microVM isolation",
      "overview": [
        "Docker Sandboxes run coding agents such as Claude Code, Codex, Copilot, Gemini CLI and others inside microVMs on your machine. Each sandbox has its own kernel and its own Docker Engine, so agents can build and run containers without touching the host’s Docker.",
        "Network access is deny-by-default and credentials are injected by a host-side proxy. A cloud version on the Docker Agentic Platform is experimental."
      ],
      "notable": [
        "Ships templates for a dozen popular coding agents.",
        "Does not require Docker Desktop on the host."
      ],
      "pros": [
        "You want to let a coding agent run unattended on your laptop safely",
        "Your agent needs to build and run Docker containers",
        "You want local sandboxes for free"
      ],
      "cons": [
        "Not open source",
        "Cloud offering is experimental, and per-unit rates are not published in the public docs"
      ],
      "verdict": "A free CLI from Docker that runs coding agents in local microVM sandboxes, suited to developers who want agents isolated from their own machine; a Docker-hosted cloud option is experimental.",
      "peers": [
        "anthropic-srt",
        "microsandbox",
        "daytona",
        "e2b"
      ],
      "sources": [
        "https://docs.docker.com/ai/sandboxes/",
        "https://docs.docker.com/ai/sandboxes/architecture.md",
        "https://docs.docker.com/ai/sandboxes/security.md",
        "https://docs.docker.com/ai/sandboxes/cloud/local-vs-cloud.md",
        "https://github.com/docker/sbx-releases",
        "https://docs.docker.com/agentic-platform/signup/",
        "https://docs.docker.com/subscription-billing/plans/docker-agentic-platform/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "anthropic-srt",
      "name": "Anthropic Sandbox Runtime",
      "company": "Anthropic",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.anthropic.com/legal/privacy"
      },
      "kind": "local",
      "tier": 3,
      "website": "https://github.com/anthropics/sandbox-runtime",
      "docs": "https://github.com/anthropics/sandbox-runtime#readme",
      "pricing_url": null,
      "github": "https://github.com/anthropics/sandbox-runtime",
      "tagline": "Open-source CLI and library (`srt`) that restricts filesystem and network access for any local process at the OS level, without a container or VM.",
      "isolation": "OS-level: macOS Seatbelt profiles; Linux bubblewrap with network namespace removed and traffic forced through host proxies; Windows support in alpha",
      "isolation_short": "OS-level (Seatbelt / bubblewrap)",
      "isolation_class": "os",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "CLI (srt)",
        "TypeScript (npm)"
      ],
      "cold_start": null,
      "cold_start_short": "Process start",
      "max_session": null,
      "persistence": "Operates on the host filesystem; writes only to explicitly allowed paths",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Deny-all by default; domain allowlists and denylists enforced by host HTTP and SOCKS5 proxies",
      "pricing": "Free and open source",
      "free_tier": "Open source",
      "compliance": null,
      "regions": null,
      "best_for": "Lightweight local sandboxing of coding agents and CLI tools",
      "overview": [
        "Sandbox Runtime is the OS-level sandbox developed for Claude Code and released as open source. It wraps any process with filesystem write restrictions and a deny-by-default network proxy, using native OS mechanisms instead of a VM.",
        "Because there is no VM, it starts instantly and works on the host filesystem — it is an OS-level process sandbox rather than a container or VM, and is labelled a beta research preview."
      ],
      "notable": [
        "Labelled “Beta Research Preview”.",
        "Reads allowed by default; writes denied by default."
      ],
      "pros": [
        "You want near-zero-overhead sandboxing for a local agent or CLI",
        "You need strict network allowlisting on a developer machine",
        "You want to wrap existing tools without containers"
      ],
      "cons": [
        "OS-level process sandbox, not a container or VM boundary",
        "Beta; configuration format may change",
        "Not a hosted or multi-tenant service"
      ],
      "verdict": "An OS-level sandbox that restricts filesystem and network access for local agents and CLI tools without a container or VM, suited to developer machines; Anthropic labels it a beta research preview.",
      "peers": [
        "docker-sandboxes",
        "microsandbox"
      ],
      "sources": [
        "https://github.com/anthropics/sandbox-runtime"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "aws-lambda-microvms",
      "name": "AWS Lambda MicroVMs",
      "company": "Amazon Web Services",
      "hq": {
        "city": "Seattle",
        "region": "Washington",
        "country": "US",
        "source": "https://aws.amazon.com/privacy/"
      },
      "kind": "hyperscaler",
      "tier": 2,
      "alternatives": true,
      "website": "https://aws.amazon.com/lambda/",
      "docs": "https://docs.aws.amazon.com/lambda/latest/dg/lambda-microvms-guide.html",
      "pricing_url": "https://aws.amazon.com/lambda/pricing/",
      "github": null,
      "tagline": "A Lambda resource type that runs a full Amazon Linux 2023 environment in a Firecracker microVM, restored from a pre-built snapshot, with suspend/resume and an HTTPS endpoint per VM.",
      "isolation": "Firecracker microVM per MicroVM (\"VM-level isolation with full operating system capabilities\"); each runs Amazon Linux 2023 and is restored from a Firecracker snapshot of the built image",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "PHP",
        "CLI",
        "REST API"
      ],
      "cold_start": "Docs claim \"near-instant launch and resume\" from snapshot; the Claude Managed Agents integration page says \"Resume from Firecracker snapshot in sub-second to single-digit seconds\"",
      "cold_start_short": "sub-second to seconds (claimed)",
      "max_session": "maximumDurationInSeconds is 1-28,800 seconds (8 hours) including suspended time; the quota page lists \"Maximum execution duration per MicroVM: 8 hours\", not adjustable",
      "max_session_short": "8 h",
      "persistence": "Suspend preserves memory and disk state (checkpointed), resume restores it; auto-suspend after a configurable idle period and auto-resume when traffic arrives; no compute charges while suspended; images are versioned Firecracker snapshots built from a Dockerfile",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not documented — sizes are CPU/memory only (0.25-4 vCPU baseline, burstable to 4x); GPUs are not mentioned in the docs",
      "networking": "Inbound via a per-VM HTTPS endpoint requiring a JWE auth token (HTTP/1.1, HTTP/2, WebSockets, gRPC, SSE; default port 8080); NO_INGRESS connector disables inbound; outbound is public internet by default or a customer VPC egress connector; PrivateLink supported",
      "pricing": "US East (N. Virginia), Arm: $0.0000276944 per vCPU-second and $0.0000036667 per GB-second while running; snapshot write $0.0038/GB (suspend), snapshot read $0.00155/GB (resume or launch); image and suspended-state storage $0.08 per GB-month (one-week minimum for images); billed per second at the baseline rate plus active use above baseline",
      "free_tier": null,
      "compliance": null,
      "regions": "14 AWS Regions as of 22 Oct 2026: launched in US East (N. Virginia, Ohio), US West (Oregon), Europe (Ireland), Asia Pacific (Tokyo); later Mumbai, Singapore, Sydney, Frankfurt, Stockholm, Hong Kong, Malaysia, Thailand and Spain",
      "best_for": "AWS-native agent sandboxes and interactive sessions that need VM isolation, an inbound endpoint and cheap idle time via suspend",
      "overview": [
        "AWS Lambda MicroVMs is a resource type in AWS Lambda, announced on 22 June 2026, for \"sandboxes for AI\", interactive development environments and multi-tenant CI. You upload a zip containing a Dockerfile and application files to S3, Lambda builds it on a managed Amazon Linux 2023 base image and captures a **Firecracker snapshot** of the fully initialised disk and memory state. Calling `run-microvm` restores a MicroVM from that snapshot, assigns it a unique HTTPS endpoint and a lifecycle of PENDING → RUNNING → SUSPENDED → TERMINATED. You drive it with the AWS CLI (`aws lambda-microvms`), the AWS SDKs (boto3, `@aws-sdk/client-lambda-microvms`, PHP and others) or the REST API.",
        "Each MicroVM is sized by baseline memory (0.5-8 GB, 0.25-4 vCPU) and can burst vertically to four times that; disk is 8-32 GB; x86_64 and ARM64 are supported. A MicroVM can live up to 8 hours in total. Idle policies suspend it after a configurable period, preserving memory and disk; traffic can auto-resume it. Pricing is per-second compute plus snapshot read/write and storage charges. AWS documents a reference pattern using MicroVMs as self-hosted sandboxes for Claude Managed Agents, where the agent loop stays with Anthropic and tool calls run inside the MicroVM. It is a managed AWS service and runs only in AWS."
      ],
      "notable": [
        "Announced 22 June 2026 in five Regions; expanded to 10 Regions in August 2026 and 14 in October 2026.",
        "Default account quota: 400 GB of memory across all RUNNING or SUSPENDED MicroVMs per Region (1,024 GB in N. Virginia, Oregon, Ohio and Tokyo), burstable to four times that; new accounts get reduced quotas.",
        "Hard limit of 8 hours per MicroVM including suspended time; RunMicrovm API rate is 5 TPS by default (adjustable)."
      ],
      "pros": [
        "You want Firecracker-isolated sandboxes inside your own AWS account with IAM, VPC egress and PrivateLink",
        "You need sessions that sit idle for long periods without paying for compute (suspend/resume with state intact)",
        "You run inbound-facing workloads (dev servers, WebSockets, gRPC) that need a per-sandbox authenticated endpoint",
        "You are building self-hosted sandboxes for Claude Managed Agents on AWS"
      ],
      "cons": [
        "Hard 8-hour maximum lifetime per MicroVM, including suspended time",
        "GPU sizes are not documented; largest baseline is 8 GB / 4 vCPU (32 GB / 16 vCPU at peak) with 32 GB disk",
        "Images must be built by Lambda from a Dockerfile on a managed base image; snapshot compatibility caveats apply (unique state generated at build is shared by every MicroVM)",
        "An exec/files SDK abstraction is not documented — you expose your own HTTP service inside the VM and talk to it through the endpoint; shell access is offered for development and troubleshooting"
      ],
      "verdict": "An AWS-native way to get Firecracker-isolated, suspendable sandboxes with a per-VM endpoint and an 8-hour maximum lifetime, driven through the AWS CLI, SDKs and your own HTTP service rather than a dedicated sandbox SDK.",
      "peers": [
        "aws-agentcore",
        "azure-dynamic-sessions",
        "gke-agent-sandbox",
        "e2b",
        "fly-machines",
        "modal"
      ],
      "sources": [
        "https://docs.aws.amazon.com/lambda/latest/dg/lambda-microvms-guide.html",
        "https://docs.aws.amazon.com/lambda/latest/dg/microvms-how-it-works.html",
        "https://docs.aws.amazon.com/lambda/latest/dg/microvms-launching.html",
        "https://docs.aws.amazon.com/lambda/latest/dg/microvms-images.html",
        "https://docs.aws.amazon.com/lambda/latest/dg/microvms-networking.html",
        "https://docs.aws.amazon.com/lambda/latest/dg/microvms-integrations-claude-managed-agents.html",
        "https://docs.aws.amazon.com/lambda/latest/dg/gettingstarted-limits.html",
        "https://aws.amazon.com/lambda/pricing/",
        "https://aws.amazon.com/about-aws/whats-new/2026/06/aws-lambda-microvms/",
        "https://aws.amazon.com/about-aws/whats-new/2026/10/lambda-microvms-4-additional-regions/",
        "https://aws.amazon.com/blogs/aws/run-isolated-sandboxes-with-full-lifecycle-control-aws-lambda-introduces-microvms/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "fly-machines",
      "name": "Fly Machines",
      "company": "Fly.io",
      "hq": {
        "city": "Chicago",
        "region": "Illinois",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/fly-io"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://fly.io/machines",
      "docs": "https://docs.fly.io/machines",
      "pricing_url": "https://docs.fly.io/about/pricing",
      "github": "https://github.com/superfly/fly-go",
      "tagline": "Fly.io's low-level API for Firecracker VMs that start and stop in under a second and can be suspended to disk; the primitive underneath Fly.io's Sprites product.",
      "isolation": "Firecracker microVM per Machine (\"Compute jobs at Fly.io are virtualized using Firecracker\"); runs on Fly.io's own hardware in 17 regions",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "Go",
        "CLI",
        "REST API"
      ],
      "cold_start": "Docs: starting a created Machine \"usually takes well under a second\"; creating one (image fetch, rootfs build) \"maybe low double digit seconds\"; resuming a suspended Machine \"takes a few hundred milliseconds\"; marketing claims \"sub-second starts and stops\"",
      "cold_start_short": "<1 s start (claimed)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Stopped Machines keep their rootfs assembled on the host for fast restart (ephemeral storage is a clean slate on each start); suspended Machines \"preserve memory and disk state\" and resume faster than stopped; attach a Fly Volume for persistent storage; volume snapshots",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "GPU Machines have been offered; the GPU docs URL now redirects to a blog post that says Fly is \"scaling back our GPU ambitions\" but \"we're not getting rid of them\"; no GPU rates appear on the pricing page",
      "networking": "Private IPv6 network per organisation (WireGuard access); public IPv4/IPv6 and Fly Proxy routing on request; static egress IPs available (`fly ips allocate-egress`); inbound services optional",
      "pricing": "Billed by the second while running; stopped Machines pay only rootfs at $0.15/GB-month. Monthly equivalents in the cheapest regions: shared-cpu-1x 256 MB $2.19, 512 MB $3.69, 1 GB $6.70; shared-cpu-2x 2 GB $13.39; performance-1x 2 GB $33.00; performance-2x 4 GB $66.00; extra RAM $6/GB-month; volumes $0.15/GB-month; snapshots $0.08/GB-month after 10 GB free; 40% off reserved compute blocks; prices vary by region",
      "free_tier": "$1 trial credit for 7 days (no card required); personal organisations with monthly bills under $5 are waived; no ongoing free tier",
      "compliance": "SOC 2 Type 2; BAA and SOC 2 report available in the Compliance Package via the dashboard Trust Center; DPAs on request",
      "regions": "17 regions: Amsterdam, Stockholm, Paris, Dallas, Secaucus, Frankfurt, São Paulo, Ashburn, Johannesburg, Los Angeles, London, Tokyo, Chicago, Singapore, San Jose, Sydney, Toronto",
      "best_for": "Teams that want raw, per-second-billed Firecracker VMs they orchestrate themselves rather than a sandbox SDK",
      "overview": [
        "Fly Machines are the compute primitive of Fly.io: \"fast-launching VMs\" created from any OCI image and controlled through the Machines REST API or the `flyctl` CLI. Each Machine is a Firecracker microVM with a lifecycle you own — `created`, `started`, `stopped`, `suspended`, `destroyed` — and the docs describe starting a created Machine in \"well under a second\". The official Go client (`fly-go`) backs `flyctl`; other languages generate clients from the OpenAPI spec. Fly.io also has a separate, higher-level product for agents, **Sprites**, which has its own API and SDKs (JavaScript/TypeScript, Python, Go, Elixir, MCP) — that is covered as `fly-sprites`; this record is the lower-level Machines API.",
        "You pick a region, a CPU kind (`shared` or `performance`) and memory (2 GB per shared CPU or 8 GB per performance CPU at most), optionally attach a Fly Volume, and are billed by the second while running; a stopped Machine costs only its rootfs storage. Machines can be stopped or suspended automatically by Fly Proxy when idle, and `fly machine run --shell` gives a throwaway interactive shell. A maximum lifetime, a dedicated exec/file SDK for agents and managed network allow-listing are not documented — flyctl offers `fly machine exec`, and the rest you build yourself on top of the VM. Fly.io holds SOC 2 Type 2."
      ],
      "notable": [
        "Fly.io states its compute is virtualised with Firecracker, \"the virtualization engine AWS developed for Lambda and Fargate\".",
        "Docs give concrete timing: started-from-stopped \"about 2 seconds or more for common apps\", resume-from-suspend \"a few hundred milliseconds\".",
        "Fly.io's agent-focused product, Sprites, is billed and documented separately from Machines (Sprites rate card: $0.0385/CPU-hour, $0.021875/GB-hour memory)."
      ],
      "pros": [
        "You want per-second-billed Firecracker VMs with full control of lifecycle, region placement and networking",
        "You need suspend-to-disk and sub-second restarts for bursty or idle-heavy workloads",
        "You are already on Fly.io and want sandboxes on the same private network as your app",
        "You run long-lived or persistent workloads (volumes, no documented session cap)"
      ],
      "cons": [
        "No dedicated sandbox SDK: file-transfer and snapshot-fork APIs for agents are not documented, and command execution is offered through flyctl (`fly machine exec`)",
        "Only an official Go client; other languages must be generated from the OpenAPI spec",
        "Machine placement can fail when a region is out of capacity, per the docs",
        "Fly says it is \"scaling back our GPU ambitions\"; the pricing page lists no GPU rates and the GPU docs URL redirects to that blog post"
      ],
      "verdict": "A per-second-billed Firecracker VM primitive for teams that want to assemble their own sandbox layer; Fly.io's separate Sprites product is its agent-focused offering.",
      "peers": [
        "fly-sprites",
        "e2b",
        "unikraft-cloud",
        "aws-lambda-microvms",
        "koyeb",
        "railway-sandboxes"
      ],
      "sources": [
        "https://fly.io/machines.md",
        "https://docs.fly.io/machines",
        "https://docs.fly.io/machines/overview",
        "https://docs.fly.io/machines/api",
        "https://docs.fly.io/machines/machine-states",
        "https://docs.fly.io/machines/guides-examples/machine-sizing",
        "https://docs.fly.io/launch/autostop-autostart",
        "https://docs.fly.io/about/pricing",
        "https://docs.fly.io/security/security-at-fly-io",
        "https://docs.fly.io/reference/regions",
        "https://fly.io/blog/wrong-about-gpu/",
        "https://fly.io/sprites/",
        "https://docs.fly.io/flyctl/machine-exec",
        "https://docs.fly.io/machines/api/machines-resource"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "unikraft-cloud",
      "name": "Unikraft Cloud",
      "company": "Unikraft",
      "hq": {
        "city": "Heidelberg",
        "region": "Baden-Württemberg",
        "country": "DE",
        "source": "https://unikraft.com/terms"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://unikraft.com",
      "docs": "https://unikraft.com/docs",
      "pricing_url": "https://unikraft.com/pricing",
      "github": null,
      "tagline": "Millisecond-start microVMs built from Dockerfiles on a modified Firecracker VMM, with stateful scale-to-zero, live forking (\"branching\") and checkpoints, plus a Sandbox SDK for agents.",
      "isolation": "Hardware-isolated microVM per instance with its own kernel, run on a \"Modified Firecracker VMM\" (KVM); sandbox egress passes through a separate \"network shield\" microVM that holds credentials",
      "isolation_short": "microVM (modified Firecracker)",
      "isolation_class": "microvm",
      "open_source": null,
      "license": null,
      "self_host": true,
      "byoc": true,
      "sdks": [
        "JavaScript/TypeScript",
        "Python",
        "Go",
        "CLI",
        "REST API"
      ],
      "cold_start": "Vendor claims \"<10ms\" cold start and scale-to-zero for all workloads, \"branch & restore in < 10 ms\", and \"Fork a live virtual computer in 10 milliseconds\"; the how-it-works page says apps are pre-initialised and snapshotted so a request to a dead-cold instance is answered \"within a single round trip\"",
      "cold_start_short": "<10 ms (claimed)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Stateful scale-to-zero (standby instances resume with memory intact); templates are snapshotted instances that preserve memory, open files and caches; branching clones a running sandbox including memory and volume state; checkpoints keep a history you can rewind to; block volumes; live migration between servers; \"autokill\" terminates long-running sandboxes",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Preview for enterprise customers only: requires the `full` (QEMU) instance type, at most one GPU per instance and you cannot choose which (example model NVIDIA Tesla T4); full VMs do not support branching or checkpointing",
      "networking": "Each instance gets a private IP and `.internal` FQDN; service groups expose a public domain `<name>.<metro>.unikraft.app`; sandbox traffic is filtered by the network-shield microVM so \"nothing reaches the internet directly\"",
      "pricing": "Hobby $0/month (2 running and 5 standby instances), Team $39/month (8 running, 50 standby), Pro $199/month (24 running, 5,000 standby); Enterprise (dedicated hosts, on-prem, BYOC) custom",
      "free_tier": "Hobby plan at $0/month (\"Free forever\"): 2 running instances, 5 standby instances, 2 vCPUs and 4 GiB memory in total",
      "compliance": "SOC 2 Type II; HIPAA-compliant; independent third-party penetration tested (stated on unikraft.com; the trust centre at trust.unikraft.com requires an access request)",
      "regions": "Default metros: Dallas, Frankfurt, San Francisco, Singapore, Washington DC; around two dozen further metros available on request",
      "best_for": "Agent fleets that need thousands of sandboxes per server, millisecond stateful wake-ups and live forking of running environments",
      "overview": [
        "Unikraft Cloud (formerly at unikraft.cloud, now unikraft.com) turns a Dockerfile into a specialised microVM image and runs it on a modified Firecracker VMM, \"based on Linux Foundation's Unikraft project\". Its pitch is density and latency: \"100K+\" microVMs per server and sub-10 ms cold starts, scale-to-zero and resume, all vendor claims. You deploy with the `unikraft` CLI, the REST API, or SDKs for JavaScript/TypeScript (`@unikraft/cloud`, which includes a `Sandbox` class with `exec`, `writeFile` and PTY-style control), Python (`unikraft-cloud`, async) and Go. Sandboxes sit behind a dedicated \"network shield\" microVM that filters egress and injects credentials so sandboxed code never sees API keys.",
        "State features are the differentiator: instances scale to zero statefully, templates are memory snapshots you clone from, **branching** forks a running sandbox (memory and volume) in milliseconds for sub-agents, and checkpoints give a rewindable history; live migration moves running VMs between hosts. GPU support exists only as an enterprise preview on QEMU-backed full VMs, which lose branching and checkpointing. Pricing is published as plans (Hobby $0, Team $39/month, Pro $199/month), with a custom-priced Enterprise plan covering dedicated hosts, on-prem and BYOC, including running as nodes in EKS, GKE or AKS. The company reports SOC 2 Type II and HIPAA compliance."
      ],
      "notable": [
        "Vendor claims \"<10ms\" cold start and \"100K+\" microVMs per server; Browser Use is cited as a customer with \"8ms cold start times\".",
        "GPU support is \"currently only available as a preview for enterprise customers\" and limited to one GPU per `full` instance.",
        "Deployment options span hosted, dedicated hosted, BYOC and on-prem, including bare metal and VMs on AWS EC2, GCP Compute Engine and Azure."
      ],
      "pros": [
        "You need very high sandbox density and millisecond resume from idle",
        "You want to fork a running agent environment (memory included) for parallel sub-agents",
        "You need a credential-holding egress proxy so sandboxed code never sees secrets",
        "You run on your own hardware or cloud and want the same stack on-prem or BYOC"
      ],
      "cons": [
        "Self-serve plans cap running instances (2 on Hobby, 8 on Team, 24 on Pro); higher limits need the custom-priced Enterprise plan",
        "GPU support is an enterprise preview on a different (QEMU) instance type without branching or checkpoints",
        "A licence or public source repository for the platform is not documented; the site says it is built on the Linux Foundation's Unikraft project",
        "Per-instance maximum sizes and session limits are not documented; plans publish total vCPU and memory quotas"
      ],
      "verdict": "A density- and latency-focused microVM platform whose stateful forking and shielded egress suit large agent fleets, sold as Hobby, Team and Pro plans with custom Enterprise pricing.",
      "peers": [
        "fly-machines",
        "e2b",
        "morph",
        "aws-lambda-microvms",
        "blaxel",
        "modal"
      ],
      "sources": [
        "https://unikraft.com/",
        "https://unikraft.com/index.md",
        "https://unikraft.com/pricing",
        "https://unikraft.com/how-it-works.md",
        "https://unikraft.com/docs",
        "https://unikraft.com/docs/use-cases/sandboxes",
        "https://unikraft.com/docs/platform/instances",
        "https://unikraft.com/docs/platform/metros",
        "https://unikraft.com/docs/sdks",
        "https://unikraft.com/blog/ai-sandboxes-at-scale.md",
        "https://unikraft.com/ai.md",
        "https://unikraft.com/customer-stories/browser-use-sandboxing"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "lightning-ai-sandboxes",
      "name": "Lightning AI Sandbox",
      "company": "Lightning AI",
      "hq": {
        "city": "New York",
        "region": "New York",
        "country": "US",
        "source": "https://lightning.ai/legal/privacy"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://lightning.ai",
      "docs": "https://lightning.ai/docs/platform/build/sandbox/",
      "pricing_url": "https://lightning.ai/pricing",
      "github": null,
      "tagline": "Isolated, programmatic Linux sandboxes on the Lightning AI platform with Python and TypeScript SDKs, a CLI, filesystem snapshots and optional Docker-in-sandbox.",
      "isolation": "Not named; docs say each sandbox runs in a \"strongly isolated environment, keeping its filesystem, network, and processes separate\" with \"dedicated compute\" and its own network stack",
      "isolation_short": "Not named (isolated Linux env)",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": true,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "CLI"
      ],
      "cold_start": "Docs claim \"Sandboxes start in milliseconds\"",
      "cold_start_short": "ms (claimed)",
      "max_session": "Configurable per sandbox (`--timeout` in milliseconds auto-stops it); no default or maximum published",
      "max_session_short": null,
      "persistence": "Persistent sandboxes \"auto-save state on stop and resume right where you left off\" (stop/start/resume); ephemeral mode available; filesystem snapshots (not running processes) can be created, listed, given a TTL and restored into a new sandbox",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not documented for Sandbox — the listed instance types are CPU only (cpu-1 1 vCPU/1 GB to cpu-16 16 vCPU/64 GB); the wider Lightning platform sells GPUs",
      "networking": "Expose ports with `--port` and get a live preview URL for dev servers; internet access and allow-listing are not described",
      "pricing": "No Sandbox-specific rates published. Platform plans: Free $0 (up to 30 starter credits), Pro $50/month (40 credits/month), Teams $140/user/month (50 credits/month), Enterprise custom; usage is billed by the second in credits",
      "free_tier": "Free plan with 5 credits on registration and 25 more on adding a card; Sandbox allowance not stated separately",
      "compliance": "SOC 2 Type II (\"with support for AWS and GCP Clouds in all regions\"); HIPAA listed for Enterprise and private deployments",
      "regions": null,
      "best_for": "Teams already on Lightning AI who want agent sandboxes next to their Studios and GPU jobs",
      "overview": [
        "Lightning AI Sandbox is the sandbox product of Lightning AI's cloud platform (the company behind PyTorch Lightning). The docs describe it as giving \"every workload its own isolated Linux environment, spun up in milliseconds and torn down when you're done\", aimed at running LLM-generated code, agent output, repository tests and live preview servers. You create sandboxes from the `lightning-sdk` Python package, the `@lightningai/sdk` TypeScript package or the `lightning sandbox` CLI, with runtimes `node24` (default), `node22`, `python313` or your own OCI image, always with sudo. A `--docker` flag provisions a Docker-enabled sandbox and `--spot` places it on spot capacity.",
        "Sandboxes can be ephemeral or persistent; persistent ones stop and resume with their state auto-saved, and filesystem snapshots can be captured with an expiry and restored into new sandboxes. Five CPU shapes run from 1 vCPU/1 GB to 16 vCPU/64 GB; no GPU shapes are listed for Sandbox. The isolation technology is not named in the docs. Lightning publishes no Sandbox-specific rate card — usage draws on the platform's per-second credit billing (Free, Pro $50/month, Teams $140/user/month). The company holds SOC 2 Type II and offers private deployments in the customer's AWS or GCP account."
      ],
      "notable": [
        "Docs claim sandboxes \"start in milliseconds\"; the CLI supports `--spot` placement and a `--docker` mode that starts a Docker daemon at boot.",
        "Platform pricing page: Pro $50/month, Teams $140/user/month; up to 30 free credits to start.",
        "Lightning states SOC 2 Type II compliance covering AWS and GCP in all regions; HIPAA is tied to Enterprise and private deployments."
      ],
      "pros": [
        "You want Python and TypeScript SDKs plus a CLI with persistent or ephemeral sandboxes on one platform",
        "You need Docker inside the sandbox or spot pricing for cheap batch runs",
        "You are already using Lightning Studios or GPU jobs and want sandboxes in the same teamspace"
      ],
      "cons": [
        "Isolation technology is not named in the docs",
        "Sandbox-specific pricing, GPU instance types, session limits and regions are not documented",
        "Networking controls (egress allow-listing) are not described"
      ],
      "verdict": "An SDK-driven sandbox for Lightning AI users with Python and TypeScript SDKs, a CLI and filesystem snapshots; the isolation technology and Sandbox-specific rates are not documented.",
      "peers": [
        "modal",
        "e2b",
        "daytona",
        "baseten-sandboxes",
        "prime-intellect-sandboxes",
        "beam"
      ],
      "sources": [
        "https://lightning.ai/docs/platform/build/sandbox/",
        "https://lightning.ai/docs/cli/stable/command-details/sandbox",
        "https://lightning.ai/docs/platform/developers/sdk.md",
        "https://lightning.ai/docs/platform/developers/sdk/sandbox.md",
        "https://lightning.ai/docs/platform/build/sandbox/examples",
        "https://lightning.ai/pricing",
        "https://lightning.ai/docs/platform/security/compliance/soc2-type2/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "prime-intellect-sandboxes",
      "name": "Prime Sandboxes",
      "company": "Prime Intellect",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.primeintellect.ai/careers/7da59a70-1e43-465d-a65b-bd495b5d5b3f"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.primeintellect.ai",
      "docs": "https://docs.primeintellect.ai/sandboxes/overview",
      "pricing_url": "https://docs.primeintellect.ai/sandboxes/overview",
      "github": "https://github.com/PrimeIntellect-ai/verifiers",
      "tagline": "Disposable, hardware-isolated microVMs that boot OCI images with Docker inside, built for agentic RL rollouts at the scale of thousands of concurrent sandboxes.",
      "isolation": "Hardware-virtualised microVM with its own guest kernel per sandbox (hypervisor not named); each boots an OCI/Docker image as its root filesystem and can run Docker, systemd and background services",
      "isolation_short": "microVM, own kernel",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "CLI",
        "REST API"
      ],
      "cold_start": "A per-sandbox boot figure is not published; docs say \"Thousands can start concurrently in seconds\" and the launch post says it can \"deploy thousands of sandboxes within seconds\"",
      "cold_start_short": null,
      "max_session": "Lifetime timeout 1 minute to unlimited (default 60 minutes); idle timeout 1 minute to 1,440 minutes (24 h), disabled by default",
      "max_session_short": "60 min default, unlimited",
      "persistence": "Pause/resume keeps memory and processes intact (paused sandboxes billed for disk only); filesystem checkpoints can be restored or forked and reach a DURABLE state; persistent named volumes can be attached and published; custom images built in the cloud",
      "snapshots": true,
      "pause_resume": true,
      "gpu": false,
      "gpu_detail": "Listed as \"coming soon\"; limits table allows 0-8 GPUs but `--gpu-count` above 0 errors until then",
      "networking": "Outbound allow/deny by domain or CIDR, updatable on a running sandbox; header secret injection into HTTPS requests; private sandbox-to-sandbox network at `<sandbox-id>.sandbox.internal` with per-port rules; public ports and tunnels expose an HTTPS URL",
      "pricing": "$0.02 per vCPU-hour, $0.0125 per GiB-hour memory, $0.0002 per GiB-hour disk (launch rates stated as valid through 22 December 2026); example 1 vCPU / 2 GiB / 5 GiB = $0.046/hour; paused sandboxes pay disk only; image and checkpoint storage $0.10 per GiB-month billed hourly, plus a one-day storage charge on upload",
      "free_tier": null,
      "compliance": null,
      "regions": null,
      "best_for": "RL environment rollouts and agent evaluations that need thousands of Docker-capable microVMs at low hourly cost",
      "overview": [
        "Prime Sandboxes is the sandbox product of Prime Intellect, the company behind the `verifiers` and `prime-rl` open-source libraries. Each sandbox is a \"hardware-virtualized microVM with its own guest kernel\" (the hypervisor is not named) that boots an OCI image as its root filesystem, so Docker, Docker Compose, systemd and kernel-dependent workloads run unmodified. You use the `prime_sandboxes` Python SDK (sync and async clients, fleets), the `prime sandbox` CLI or the API; the `verifiers` library ships `SandboxEnv` and `PythonEnv` classes that use them for RL environments.",
        "Sandboxes range from 1 to 16 vCPU, 128 MiB to 64 GiB memory and 2 to 128 GiB disk; the default is 1 vCPU / 1 GiB / 5 GiB with a 60-minute timeout, and lifetimes can be unlimited. Pause/resume preserves memory and processes, filesystem checkpoints can be forked, and outbound networking is governed by domain or CIDR allow/deny rules with secret injection into HTTPS headers. Pricing is purely usage-based ($0.02/vCPU-hour, $0.0125/GiB-hour). New accounts default to 1,024 active sandboxes and 102,400 creations per hour. The product reached general availability on 23 September 2026; GPU sandboxes are \"coming soon\"."
      ],
      "notable": [
        "General availability announced 23 September 2026; the launch post cites dashboards showing \"more than 20,000 concurrent sandboxes\".",
        "Launch pricing ($0.02/vCPU-hour, $0.0125/GiB-hour memory) is stated as valid through 22 December 2026.",
        "Default account limits: 1,024 active sandboxes, 4,096 vCPUs, 4,096 GiB memory, 32,768 GiB storage."
      ],
      "pros": [
        "You run large RL or evaluation fleets and want thousands of concurrent microVMs with Docker inside",
        "You need fine-grained egress rules and header-level secret injection so agents never hold credentials",
        "You want pause/resume with memory intact and forkable filesystem checkpoints at low hourly rates",
        "You already use `verifiers` or `prime-rl` for environments"
      ],
      "cons": [
        "Hypervisor, regions and compliance are not documented",
        "The documented SDK is Python (plus CLI and API); a TypeScript SDK is not documented",
        "GPU sandboxes are not yet available",
        "A free tier is not published"
      ],
      "verdict": "A low-cost, RL-oriented microVM sandbox with strong networking and checkpoint features, best for Python-centric teams running rollouts at scale.",
      "peers": [
        "e2b",
        "daytona",
        "modal",
        "lightning-ai-sandboxes",
        "novita-sandbox",
        "morph"
      ],
      "sources": [
        "https://www.primeintellect.ai",
        "https://docs.primeintellect.ai/sandboxes/overview",
        "https://docs.primeintellect.ai/sandboxes/cli",
        "https://www.primeintellect.ai/blog/sandboxes"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "val-town",
      "name": "Val Town",
      "company": "Val Town",
      "hq": {
        "city": "Brooklyn",
        "region": "New York",
        "country": "US",
        "source": "https://www.val.town/careers"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "status": "winding-down",
      "status_note": "Val Town announced on **November 10, 2026** that it is winding down: new sign-ups and plan upgrades closed that day, code becomes read-only on December 10, 2026 and remaining services end on **January 10, 2027**. Billing was paused on the day of the announcement.",
      "website": "https://www.val.town",
      "docs": "https://docs.val.town",
      "pricing_url": "https://www.val.town/pricing",
      "github": "https://github.com/val-town",
      "tagline": "A hosted TypeScript runtime where \"vals\" (HTTP endpoints, crons, email handlers) deploy to live URLs and run under Deno permissions with no filesystem or subprocess access; Val Town announced on 10 November 2026 that it is winding down, with services ending on 10 January 2027.",
      "isolation": "Deno process per val with a restricted permission set: `--allow-net` and env (secrets) on; filesystem read/write, subprocess execution (`--allow-run`), FFI and high-resolution time denied; the 2024 engineering blog describes \"temporary Deno subprocesses\" from a pre-initialised pool; no container or VM layer is documented",
      "isolation_short": "Deno process sandbox",
      "isolation_class": "os",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": "Homepage claims vals are \"Deployed to the web in 100ms\"",
      "cold_start_short": "100 ms deploy (claimed)",
      "max_session": "Wall-clock time per run: 1 minute (Free), 10 minutes (Pro); Business and Enterprise not listed",
      "max_session_short": "1 min / 10 min (Pro)",
      "persistence": "Vals are persistent deployed code with version history and branches; no filesystem — state goes to the built-in SQLite database or blob storage (10 MB SQLite free, up to 1 GB paid); in-memory state persists only until a val is restarted",
      "snapshots": false,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented — the runtime is JavaScript/TypeScript only",
      "networking": "Outbound HTTP allowed (`--allow-net`); every HTTP val gets a public live URL; custom domains on paid plans",
      "pricing": "Free $0; Pro $21/month (1-minute crons, 10-minute runs, 2,000 private vals, 10 custom domains, 1M runs/day); Business from $167/month (5M runs/day, unlimited custom domains); Enterprise custom; plan upgrades closed and billing was paused on 10 November 2026",
      "free_tier": "Free plan: unlimited public vals, 15-minute cron interval, 1-minute runs, 100,000 runs/day, 3-day log retention, no private vals; new sign-ups closed on 10 November 2026",
      "compliance": "SOC 2 is listed under the Enterprise plan on the pricing page",
      "regions": null,
      "best_for": "Existing Val Town users planning a migration before services end on 10 January 2027; new sign-ups are closed",
      "overview": [
        "Val Town describes itself as \"Instant deployments for you & your agents\": a hosted platform where developers and AI agents write TypeScript \"vals\" that run as HTTP endpoints, scheduled crons, email handlers or scripts and go live at a public URL. Code runs on Deno (currently Deno 2.9.x with TypeScript 5.9) with a locked-down permission set — network and secrets allowed, filesystem, subprocesses, FFI and high-resolution timers denied — which Val Town documents as its sandbox. A 2024 engineering post describes the fourth-generation runtime as pooled Deno subprocesses of a Node.js server; whether a container or VM sits beneath that is not documented.",
        "Agents reach it through the official MCP server, the `@valtown/vt` CLI (clone, push, pull, watch, create, remix), the `@valtown/sdk` TypeScript library and a REST API, with setup guides for Claude Code, Codex, Cursor and others. State lives in a per-user SQLite database and S3-style blob storage rather than a filesystem. Limits are plan-based: 1-minute runs and 15-minute crons on Free, 10-minute runs and 1-minute crons on Pro ($21/month). It is a deployment runtime rather than a general shell sandbox: subprocess and executable execution and general filesystem access are denied, and GPU access is not documented. On 10 November 2026 Val Town announced that it is winding down: new sign-ups and plan upgrades closed that day, code becomes read-only on 10 December 2026 and remaining services end on 10 January 2027."
      ],
      "notable": [
        "Pro plan is $21/month with 10-minute wall-clock runs and 1,000,000 runs per day; Business starts at $167/month.",
        "Runtime limits per val: 1,000 files and 100 branches; binary files are not supported.",
        "Val Town's own docs recommend the Val Town plugin or MCP server over the `vt` CLI for agents because the MCP server has more tools."
      ],
      "pros": [
        "You want agents to deploy webhooks, crons, small APIs or MCP servers to a live URL in one step",
        "You need a TypeScript-only runtime with managed secrets, SQLite and blob storage built in"
      ],
      "cons": [
        "Winding down: new sign-ups closed on 10 November 2026 and services end on 10 January 2027, per Val Town's announcement",
        "No general filesystem access, subprocesses or executables; TypeScript/JavaScript only; run limits of 1 minute (Free) and 10 minutes (Pro)",
        "Isolation below the Deno permission layer is not documented; compliance information is limited to a SOC 2 entry under the Enterprise plan on the pricing page"
      ],
      "verdict": "A hosted deployment runtime for agent-written TypeScript services such as webhooks and MCP servers, without shell or general filesystem access, which Val Town is winding down by 10 January 2027.",
      "peers": [
        "deno-sandbox",
        "cloudflare-sandbox",
        "vercel-sandbox",
        "freestyle",
        "webcontainers"
      ],
      "sources": [
        "https://www.val.town",
        "https://www.val.town/pricing",
        "https://docs.val.town/reference/runtime",
        "https://docs.val.town/troubleshooting/permission-errors",
        "https://docs.val.town/vals/limitations",
        "https://docs.val.town/reference/api",
        "https://sdk.val.town/",
        "https://github.com/val-town/vt",
        "https://blog.val.town/first-four-val-town-runtimes",
        "https://blog.val.town/val-town-town",
        "https://blog.val.town/winding-down",
        "https://docs.val.town/migration"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "control-plane-sandboxes",
      "name": "Control Plane Sandboxes",
      "company": "Control Plane",
      "hq": {
        "region": "California",
        "country": "US",
        "source": "https://controlplane.com/company"
      },
      "tier": 2,
      "kind": "hosted",
      "alternatives": true,
      "website": "https://controlplane.com/platform/ai-agents",
      "docs": "https://docs.controlplane.com/guides/sandbox/getting-started",
      "pricing_url": "https://controlplane.com/pricing",
      "github": null,
      "tagline": "Persistent, suspendable development sandboxes for coding agents inside a Control Plane \"virtual cloud\", with a browser IDE, SSH/VS Code/Cursor access and identity-based access to private services.",
      "isolation": "Container on Control Plane-managed Kubernetes (platform security docs: cgroups, namespaces, restricted syscalls and capabilities); the sandbox docs do not name a hypervisor or microVM runtime, although a Control Plane blog post of 26 October 2026 says sandboxes run in dedicated Kata Containers microVMs",
      "isolation_short": "Kata microVM (per vendor blog)",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": true,
      "byoc": true,
      "sdks": [
        "CLI"
      ],
      "cold_start": "Docs: a new sandbox \"becomes ready in ~60–90 seconds\"",
      "cold_start_short": "60-90 s",
      "max_session": null,
      "max_session_short": null,
      "persistence": "`/root` is mounted from a persistent volumeset, so repos, credentials, IDE settings and uncommitted changes survive suspend/resume and restarts; suspend scales the workload to zero (storage billed only); delete removes the volumeset",
      "snapshots": null,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not documented for sandboxes; the platform sells GPU compute (T4 $0.35/hour to B200 $7.93/hour) for workloads",
      "networking": "Each sandbox has a password-protected browser IDE (`/_ide/`), browser terminal (`/_term/`) and a shareable public app URL (default port 8080); reaches private services via Universal Cloud Identity and Cloud Wormhole so \"the agent never handles a key\"",
      "pricing": "Sandbox-specific rates are not published; platform usage pricing is $0.06205654 per millicore-month vCPU, $0.00706330 per MB-month memory, General Purpose SSD $0.21224989 per GB-month, egress $0.12/GB; suspended sandboxes bill storage only",
      "free_tier": null,
      "compliance": "SOC 2 Type II, PCI DSS Level 1, HIPAA and GDPR compliant (per docs)",
      "regions": "Any location a GVC (Global Virtual Cloud) spans: AWS, GCP, Azure, OCI regions, other clouds, or your own Kubernetes clusters (BYOK)",
      "best_for": "Enterprises on Control Plane that want long-lived agent dev environments with identity-based access to their private infrastructure",
      "overview": [
        "Control Plane is a multi-cloud \"virtual cloud\" platform that orchestrates hardened Kubernetes clusters across AWS, GCP, Azure, OCI and customer hardware. Its **Sandboxes** feature (a documented product with its own docs section and a `cpln sandbox connect` CLI command added in CLI v3.12.0, July 2026) gives AI coding agents \"an isolated environment with persistent storage\" inside your virtual cloud. You create one from the console by picking an image (templates, composed toolboxes or your own OCI image), a size and an app port; it is ready in roughly 60-90 seconds and exposes a browser IDE (code-server), a browser terminal (ttyd with tmux) and a public URL, or you connect from VS Code, Cursor or SSH.",
        "The home directory lives on a persistent volumeset, so suspending a sandbox scales it to zero and bills only storage while keeping repos and credentials; resuming picks up where it left off. Private databases and APIs are reached through Control Plane's identity system rather than keys in the agent's context. The platform security docs describe container isolation (cgroups, namespaces, restricted syscalls and capabilities) and the sandbox docs do not name a microVM runtime, while a Control Plane blog post of 26 October 2026 says sandboxes run in dedicated Kata Containers microVMs. A sandbox SDK is not documented — the sandbox docs cover the console and CLI — and sandbox-specific pricing is not published beyond the platform's per-millicore and per-megabyte rates. Control Plane reports SOC 2 Type II, PCI DSS Level 1, HIPAA and GDPR compliance."
      ],
      "notable": [
        "`cpln sandbox connect` shipped in CLI v3.12.0 (3 July 2026); v3.13.0 switched SSH to a `devuser` account for de-rooted images.",
        "Docs quote readiness in \"~60–90 seconds\" and suspend billing as \"Storage only\".",
        "Platform compliance: SOC 2 Type II, PCI DSS Level 1, HIPAA, GDPR."
      ],
      "pros": [
        "You already run on Control Plane and want agent workspaces with identity-based access to private services",
        "You need persistent, suspendable environments with IDE, terminal and SSH access",
        "You want the sandbox to run in your own clouds or Kubernetes clusters under one control plane"
      ],
      "cons": [
        "A sandbox SDK is not documented; the sandbox docs describe creation through the console and connection through the CLI",
        "The sandbox docs do not name the isolation runtime; a Control Plane blog post says Kata Containers microVMs",
        "A new sandbox takes about 60-90 seconds to become ready, per the docs",
        "Sandbox-specific pricing, limits and idle-timeout settings are not covered in the sandbox docs"
      ],
      "verdict": "A persistent dev-environment style sandbox for enterprises standardised on Control Plane, created from the console and reached through a browser IDE, SSH or the CLI.",
      "peers": [
        "northflank",
        "koyeb",
        "railway-sandboxes",
        "coder",
        "ona-gitpod",
        "github-codespaces"
      ],
      "sources": [
        "https://controlplane.com/platform/ai-agents",
        "https://controlplane.com/pricing",
        "https://docs.controlplane.com/whatis",
        "https://docs.controlplane.com/guides/sandbox/getting-started",
        "https://docs.controlplane.com/guides/sandbox/container-internals",
        "https://docs.controlplane.com/cli-reference/release-notes",
        "https://docs.controlplane.com/core/security",
        "https://controlplane.com/blog/post/best-ai-agent-sandboxes-2026"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "ona-gitpod",
      "name": "Ona (formerly Gitpod)",
      "company": "Ona (part of OpenAI)",
      "hq": {
        "city": "Kiel",
        "region": "Schleswig-Holstein",
        "country": "DE",
        "source": "https://ona.com/legal/imprint"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://ona.com",
      "docs": "https://ona.com/docs",
      "pricing_url": null,
      "github": "https://github.com/gitpod-io/gitpod-sdk-go",
      "tagline": "Enterprise cloud development environments provisioned as isolated VMs (EC2 or Compute Engine) in your own VPC from a Dev Container config, with SDKs, a CLI and Codex-based background agents; hosted Ona Cloud was retired on 31 October 2026.",
      "isolation": "Each environment is \"an isolated VM with dedicated compute, storage, and networking\" — an EC2 instance on AWS runners or a Compute Engine instance on GCP runners — created from a Dev Container configuration; runners live in your VPC and the Ona-hosted management plane never sees source code or secrets",
      "isolation_short": "Dedicated cloud VM",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": true,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "CLI",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Auto-stop after a configurable inactivity timeout (`--inactivity-timeout`); no fixed maximum published",
      "max_session_short": null,
      "persistence": "Stop an environment \"to preserve its state\" and start it again; delete removes it permanently; environment classes set instance type and disk (default classes from m6i.large with 45 GB up to g5.4xlarge with 300 GB); warm pools (Enterprise)",
      "snapshots": null,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Default AWS environment classes include GPU classes on g5.4xlarge instances with 300 GB disk",
      "networking": "Environments run in your VPC; subnets can stay private with outbound through NAT or proxy; \"kernel-level policy enforcement\" and scoped credentials advertised as guardrails",
      "pricing": "Contact sales (Enterprise); Ona Cloud, whose usage-based Core tier started at $20/month, was shut down on 31 October 2026",
      "free_tier": null,
      "compliance": "SOC 2 and GDPR logos with a trust portal at trust.ona.com; details not specified on the pages read",
      "regions": "Wherever you deploy runners: any AWS or GCP region in your own account",
      "best_for": "Enterprises that want VM-isolated, Dev Container-based environments for developers and background agents inside their own AWS or GCP VPC",
      "overview": [
        "Ona is the company formerly known as Gitpod, re-founded under the new name on 2 September 2025 and now part of OpenAI, which it joined as part of the Codex team (the transaction closed on 10 August 2026). The product has three parts: **Environments**, API-first cloud development environments created from a `devcontainer.json`; **Agents**, now the Codex Agent (the earlier Ona Agent was withdrawn from Ona Cloud in August 2026 and is no longer listed in the docs); and **Guardrails** for policy and audit. Each environment is a dedicated VM — an EC2 instance for AWS runners or a Compute Engine instance for GCP runners — and the runner that provisions it lives in your VPC, while Ona hosts only the management plane. Environments are managed from the dashboard, the `ona` CLI (`environment create`, `exec`, `ssh`, `stop`, `delete`, JSON output for agents) and SDKs for Python (`gitpod-sdk`), TypeScript (`@gitpod/sdk`) and Go.",
        "The hosted compute option, Ona Cloud, was wound down on 31 October 2026, and new sign-ups had already closed when this was announced on 13 October; the docs now state \"Enterprise organizations can continue using runners in their own infrastructure\" on AWS or GCP. Pricing is therefore contact-sales. Environments stop and resume with state preserved, auto-stop on inactivity, and can use GPU instance classes. Gitpod Classic (the earlier product) was sunset for pay-as-you-go users on 15 October 2025; Ona said that deadline did not apply to Gitpod Enterprise customers."
      ],
      "notable": [
        "Rebranded from Gitpod to Ona on 2 September 2025; agreed to join OpenAI on 11 June 2026, and the transaction closed on 10 August 2026.",
        "Ona Cloud (hosted compute) shut down on 31 October 2026; Ona said it would refund October payments and unused purchased credits, and new sign-ups had already closed.",
        "Runners are deployed as ECS services in your AWS VPC or in your GCP project; environments are EC2 or Compute Engine instances."
      ],
      "pros": [
        "You need dedicated-VM environments that never leave your own AWS or GCP account",
        "You want Dev Container-defined, reproducible workspaces for both developers and background agents",
        "You are an enterprise standardising on Codex-based background agents with audit and policy controls"
      ],
      "cons": [
        "Hosted Ona Cloud is no longer available; remaining use is on runners in your own AWS or GCP account, with pricing via contact sales",
        "The documented built-in agent is the Codex Agent; the earlier Ona Agent is no longer listed in the docs",
        "Startup figures and session caps are not documented; the public trust page lists SOC 2 and GDPR without stating the report type",
        "Ona is now part of OpenAI, having joined as part of the Codex team"
      ],
      "verdict": "A VPC-native VM environment platform for enterprises that run Dev Container workspaces and Codex-based agents in their own AWS or GCP account, offered without the hosted Ona Cloud since 31 October 2026.",
      "peers": [
        "github-codespaces",
        "coder",
        "daytona",
        "control-plane-sandboxes",
        "codesandbox"
      ],
      "sources": [
        "https://ona.com",
        "https://ona.com/pricing",
        "https://ona.com/docs/ona/runners/overview",
        "https://ona.com/docs/ona/understanding/how-ona-works",
        "https://ona.com/docs/ona/environments/agent-environments",
        "https://ona.com/docs/ona/integrations/sdk",
        "https://ona.com/docs/ona/runners/ona-cloud",
        "https://ona.com/docs/ona/billing/overview",
        "https://ona.com/docs/ona/agents/overview",
        "https://ona.com/stories/gitpod-is-now-ona",
        "https://ona.com/stories/ona-joins-openai",
        "https://ona.com/stories/ona-core-shutdown",
        "https://ona.com/stories/anthropic-model-access-in-ona",
        "https://ona.com/stories/gitpod-classic-payg-sunset",
        "https://ona.com/stories/migrating-from-classic-to-ona",
        "https://trust.ona.com"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "github-codespaces",
      "name": "GitHub Codespaces",
      "company": "GitHub (Microsoft)",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://docs.github.com/en/site-policy/privacy-policies/github-general-privacy-statement"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/features/codespaces",
      "docs": "https://docs.github.com/en/codespaces",
      "pricing_url": "https://docs.github.com/en/billing/managing-billing-for-your-products/managing-billing-for-github-codespaces/about-billing-for-github-codespaces",
      "github": null,
      "tagline": "GitHub's cloud development environments: a Dev Container running on a freshly built, dedicated VM per codespace, created from the web, `gh` CLI or REST API, billed per core-hour with a free monthly quota.",
      "isolation": "Docker container on a dedicated virtual machine — \"Each codespace is hosted on its own newly-built virtual machine (VM)\" with its own virtual network; two codespaces never share a VM",
      "isolation_short": "Dedicated VM per codespace",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "CLI",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Stops after 30 minutes of inactivity by default (customisable); a fixed maximum lifetime of 12 hours even while in use; stopped codespaces are deleted after 30 days of inactivity by default (customisable)",
      "max_session_short": "12 h, 30 min idle",
      "persistence": "Stopping preserves saved changes and terminal history; /workspaces survives rebuilds; full rebuilds clear everything outside /workspaces; retained for 30 days by default",
      "snapshots": null,
      "pause_resume": true,
      "gpu": false,
      "gpu_detail": "No — machine types listed on the billing page are 2- to 32-core CPU machines only",
      "networking": "Firewalls block inbound internet connections and inter-codespace traffic; forwarded ports default to private (authenticated) and can be made organisation-visible or public",
      "pricing": "2-core $0.18/hour, 4-core $0.36/hour, 8-core $0.72/hour, 16-core $1.44/hour, 32-core $2.88/hour; storage $0.07 per GB-month; prebuilds consume Actions compute and storage",
      "free_tier": "Personal accounts: 120 core-hours and 15 GB-month per month on GitHub Free, 180 core-hours and 20 GB-month on GitHub Pro; organisations and enterprises have no included quota",
      "compliance": null,
      "regions": "Set automatically from your location or chosen manually in user settings; the region list is shown in the settings dropdown rather than documented",
      "best_for": "Developer-driven workspaces tied to GitHub repositories, where an agent assists a human inside a Dev Container",
      "overview": [
        "GitHub Codespaces is a cloud development environment attached to a repository or template. Each codespace is a Docker container built from the repo's `devcontainer.json` (or a default Ubuntu image) that runs on \"its own newly-built virtual machine\" with its own virtual network, so isolation is at the VM level. You open it in the browser, VS Code or over SSH, and create or control it from the web UI, the `gh codespace` CLI (create, ssh, stop, delete, ports, cp) or the REST API (`POST /repos/{owner}/{repo}/codespaces`, start, stop, delete). Machine types run from 2 cores / 8 GB / 32 GB to 32 cores / 128 GB / 128 GB.",
        "Lifecycle is time-limited: a codespace stops after 30 minutes idle by default, has a hard 12-hour maximum lifetime per session, and stopped codespaces are retained for 30 days before deletion. Saved files persist across stops. Billing is per core-hour ($0.18/hour for 2 cores) plus $0.07 per GB-month of storage, with free monthly quotas for personal accounts. Snapshot and fork APIs and a startup figure are not documented, and no GPU machine types are listed. GitHub's Copilot coding agent is documented separately and is not covered here."
      ],
      "notable": [
        "Personal GitHub Free accounts include 120 core-hours and 15 GB-month of Codespaces usage per month; Pro includes 180 core-hours and 20 GB-month.",
        "Limits: 12-hour maximum lifetime, 30-minute default idle timeout, 30-day default retention.",
        "Compute is billed per core at $0.09 per core-hour (2-core $0.18/hour up to 32-core $2.88/hour)."
      ],
      "pros": [
        "You want repo-linked Dev Container workspaces on dedicated VMs with a REST API and CLI",
        "You need per-user isolation, private port forwarding and GitHub secrets without running infrastructure",
        "You are a personal account that can use the free monthly core-hour quota"
      ],
      "cons": [
        "12-hour maximum lifetime per codespace session; 30-minute default idle timeout and 30-day default retention",
        "Snapshot, fork and memory-preserving pause APIs are not documented; no GPU machine types are listed",
        "Hosted by GitHub, with no documented option to run in your own cloud account; organisation and enterprise plans include no free Codespaces quota"
      ],
      "verdict": "A VM-isolated, repository-linked workspace hosted by GitHub, suited to developers and their assistants working inside a Dev Container within a 12-hour session lifetime.",
      "peers": [
        "ona-gitpod",
        "coder",
        "codesandbox",
        "daytona",
        "control-plane-sandboxes"
      ],
      "sources": [
        "https://docs.github.com/en/codespaces/about-codespaces/what-are-codespaces",
        "https://docs.github.com/en/codespaces/about-codespaces/the-codespace-lifecycle",
        "https://docs.github.com/en/codespaces/reference/security-in-github-codespaces",
        "https://docs.github.com/en/billing/managing-billing-for-your-products/managing-billing-for-github-codespaces/about-billing-for-github-codespaces",
        "https://docs.github.com/en/codespaces/setting-your-user-preferences/setting-your-default-region-for-github-codespaces",
        "https://docs.github.com/en/rest/codespaces/codespaces",
        "https://docs.github.com/en/codespaces/developing-in-a-codespace/using-github-codespaces-with-github-cli",
        "https://github.blog/changelog/2025-08-01-upcoming-deprecation-of-gpu-machine-type-in-codespaces/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "coder",
      "name": "Coder",
      "company": "Coder Technologies",
      "hq": {
        "city": "Austin",
        "region": "Texas",
        "country": "US",
        "source": "https://coder.com/legal/privacy-policy"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://coder.com",
      "docs": "https://coder.com/docs",
      "pricing_url": "https://coder.com/pricing",
      "github": "https://github.com/coder/coder",
      "tagline": "Self-hosted cloud development environments defined in Terraform (VMs, Kubernetes pods or Docker containers) for developers and coding agents, with a control-plane-hosted agent loop so workspaces hold no model credentials.",
      "isolation": "Whatever your Terraform template provisions — \"EC2 VMs, Kubernetes Pods, Docker Containers, etc.\" — connected to the control plane over a WireGuard tunnel; isolation strength is set by the template, not by Coder",
      "isolation_short": "Template-defined (VM/pod/container)",
      "isolation_class": "mixed",
      "open_source": true,
      "license": "AGPL-3.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Workspaces \"shut down automatically when idle\"; autostop schedules are template-defined, no fixed maximum",
      "max_session_short": null,
      "persistence": "Workspaces are long-lived resources that stop and start; persistent disks and home volumes depend on the template (for example Kubernetes PVCs or EC2 volumes)",
      "snapshots": null,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Depends on the infrastructure your templates target; Coder itself does not provision GPUs",
      "networking": "Workspaces connect outbound to the control plane via WireGuard/Tailnet with no inbound ports; templates can restrict egress, and the docs recommend agent templates with stricter egress since the agent loop needs no LLM access from the workspace; Premium Agent Firewall adds process-level network and command policies",
      "pricing": "Community edition free (unlimited self-hosted workspaces, up to 5 active Coder Agents, 1 platform integration); Premium priced annually per user (figure not published); AI Premium custom pricing with unlimited agents and headless API-triggered agent jobs",
      "free_tier": "Free to self-host (Community edition)",
      "compliance": null,
      "regions": "Anywhere you host it: AWS, Azure, GCP, Kubernetes, Hetzner, OVHcloud, VMware or a private data centre",
      "best_for": "Organisations that want to run coding agents in self-hosted workspaces on their own clusters and clouds with audit and governance",
      "overview": [
        "Coder is an AGPL-3.0 open-source platform (17.3k GitHub stars) for self-hosted development environments. Admins write Terraform templates that provision a workspace as an EC2 VM, a Kubernetes pod, a Docker container or any other resource; developers create workspaces from those templates in the UI or with `coder create`, connect over a WireGuard tunnel from VS Code, JetBrains, Cursor or the web terminal, and workspaces shut down automatically when idle. The control plane must be hosted by you in a private data centre or cloud. Because the infrastructure is yours, isolation is whatever the template gives you — a VM, a pod or a container.",
        "For agents, **Coder Agents** (the successor to Coder Tasks, which was deprecated in v2.34 and removed) runs the agent loop in the control plane: the LLM provider (Anthropic, OpenAI, Google, Azure, Bedrock or any OpenAI-compatible endpoint) is called from the control plane and only tool calls traverse the existing workspace tunnel, so \"the workspace itself has no knowledge of AI\" and holds no credentials. Terminal agents such as Claude Code or Codex can also be installed into templates via registry modules, and Agent Relay (beta) lets Cursor Cloud Agents and Claude Code drive self-hosted workspaces. The Community edition is free with up to five active Coder Agents; Premium (per user, annual, price unpublished) adds RBAC, audit logging, AI Gateway and Agent Firewall; AI Premium adds unlimited headless agents triggered via API."
      ],
      "notable": [
        "Licensed AGPL-3.0 with a separate enterprise licence file; 17.3k stars on GitHub.",
        "Coder Tasks was deprecated in v2.34 and its API and CLI removed; Coder Agents and the Chats API (`/api/v2/chats`) replace it.",
        "Community edition is free with unlimited self-hosted workspaces and up to 5 active Coder Agents."
      ],
      "pros": [
        "You want agent workspaces on infrastructure you already run (Kubernetes, EC2, Docker) under an open-source control plane",
        "You need the agent loop and model credentials kept out of the workspace, with network-isolated agent templates",
        "You run both developers and agents side by side with RBAC, audit logs and policy enforcement"
      ],
      "cons": [
        "Not a hosted service: you operate the control plane, provisioners and the infrastructure behind every workspace",
        "Isolation, startup time and persistence all depend on the templates you write",
        "Premium pricing is not published; active Coder Agents are capped at 5 below the AI Premium tier",
        "Coder Tasks was removed, so older agent integrations need migrating to Coder Agents"
      ],
      "verdict": "An open-source, self-hosted workspace platform for agents and developers, with isolation determined by the Terraform templates you write.",
      "peers": [
        "ona-gitpod",
        "github-codespaces",
        "control-plane-sandboxes",
        "daytona",
        "openhands-runtime"
      ],
      "sources": [
        "https://coder.com",
        "https://coder.com/pricing",
        "https://github.com/coder/coder",
        "https://coder.com/docs/ai-coder",
        "https://coder.com/docs/ai-coder/agents/",
        "https://coder.com/docs/ai-coder/agents/architecture",
        "https://coder.com/docs/ai-coder/agents/tasks-to-chats-migration",
        "https://coder.com/docs/about/architecture"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "webcontainers",
      "name": "WebContainers",
      "company": "StackBlitz",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://stackblitz.com/privacy-policy"
      },
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://webcontainers.io",
      "docs": "https://webcontainers.io/guides/introduction",
      "pricing_url": "https://webcontainers.io/enterprise",
      "github": null,
      "tagline": "StackBlitz's browser-based Node.js runtime: an operating system in WebAssembly that boots in a browser tab, runs npm/pnpm/yarn and dev servers client-side, and is free for open source with a commercial licence for production use.",
      "isolation": "Browser tab sandbox — \"everything is contained in a browser tab\"; the runtime is WebAssembly executing inside the browser's security model and requires cross-origin isolation (COOP/COEP headers) for SharedArrayBuffer",
      "isolation_short": "Browser tab (WebAssembly)",
      "isolation_class": "wasm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript"
      ],
      "cold_start": "Site claims a dev environment can start \"in milliseconds\"; the API reference notes \"Booting WebContainer is an expensive operation\"",
      "cold_start_short": "ms (claimed)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Filesystem lives in the browser tab; `mount()` loads a file tree or a binary snapshot and `export()` writes the filesystem out as JSON, binary or zip; persistence across page reloads is not documented",
      "snapshots": null,
      "pause_resume": null,
      "gpu": false,
      "gpu_detail": "No — WebAssembly Node.js runtime in the browser",
      "networking": "Runs inside the user's browser; the host page must send `Cross-Origin-Embedder-Policy: require-corp` (or `credentialless`) and `Cross-Origin-Opener-Policy: same-origin`; third-party asset blocking in Brave/Firefox can break it",
      "pricing": "\"Always free for Open Source\"; \"Licensing is required for production usage of the API in a commercial, for-profit setting\" (prototypes and POCs exempt), with pricing on request; Enterprise adds self-hosted/on-prem/VPC installation, private npm registries, SAML SSO and priority support",
      "free_tier": "Free for open source and for prototypes/POCs",
      "compliance": null,
      "regions": "Not applicable — executes in the end user's browser",
      "best_for": "In-browser coding products and agent UIs that need to run and preview Node.js code on the user's machine with zero server compute",
      "overview": [
        "WebContainers are StackBlitz's \"browser-based runtime for executing Node.js applications\": a WebAssembly-based operating system that boots inside a browser tab, runs the native `npm`, `pnpm` and `yarn`, and serves dev servers to an in-page preview. The `@webcontainer/api` package exposes `WebContainer.boot()`, `fs`, `mount()` and `spawn()`; only one instance can boot per page, and the host page must be cross-origin isolated (COOP/COEP headers). Chromium browsers are fully supported, Safari 16.4+ is in beta and Firefox is alpha. StackBlitz markets it for AI as a \"secure agent runtime\" where agents \"install packages, write files, and execute code\" with \"zero compute costs, and zero virtual machines\".",
        "Isolation is the browser's own sandbox — the code runs on the end user's CPU, so nothing touches a server — which makes it a client-side execution environment rather than a cloud sandbox. Per the docs it \"can only execute languages that are natively supported on the Web, including JavaScript and WebAssembly\", and native addons cannot run unless they are compiled to WebAssembly. On licensing, it is \"Always free for Open Source\" and for prototypes, but production use in a for-profit setting requires a commercial licence (configured with `configureAPIKey`), priced on request; Enterprise adds self-hosted/VPC installation and SAML SSO."
      ],
      "notable": [
        "Licensing: free for open source and prototypes; production commercial use requires an API key and licence; pricing on request.",
        "Browser support: Chromium fully supported, Safari \"in beta\" from 16.4 TP, Firefox alpha; cross-origin isolation headers are mandatory.",
        "Only a single WebContainer instance can be booted concurrently per page, per the API reference."
      ],
      "pros": [
        "You want agent-generated Node.js code to run and preview on the user's machine with no server compute",
        "You are building an in-browser IDE, tutorial or generative-UI product",
        "You are an open-source project or prototyping and can use it without a licence"
      ],
      "cons": [
        "Runs only languages natively supported on the web (JavaScript and WebAssembly); native addons cannot run unless compiled to WebAssembly",
        "Runs in the user's browser tab: persistence beyond the session and server-side orchestration are not documented, and only a single instance can be booted concurrently",
        "Commercial production use requires a licence, with pricing on request; full support is limited to Chromium-based browsers and cross-origin isolation headers are required"
      ],
      "verdict": "An in-browser Node.js runtime from StackBlitz for running and previewing agent-written code client-side in a browser tab rather than on a server.",
      "peers": [
        "pyodide",
        "codesandbox",
        "val-town",
        "deno-sandbox",
        "wasmtime"
      ],
      "sources": [
        "https://webcontainers.io",
        "https://webcontainers.io/guides/introduction",
        "https://webcontainers.io/guides/browser-support",
        "https://webcontainers.io/guides/configuring-headers",
        "https://webcontainers.io/api",
        "https://webcontainers.io/enterprise",
        "https://webcontainers.io/ai",
        "https://webcontainers.io/guides/troubleshooting"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "anthropic-code-execution",
      "name": "Claude code execution tool",
      "company": "Anthropic",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.anthropic.com/legal/privacy"
      },
      "kind": "hyperscaler",
      "tier": 2,
      "alternatives": true,
      "website": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool",
      "docs": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool",
      "pricing_url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool#usage-and-pricing",
      "github": null,
      "tagline": "Server-side tool in the Claude API that runs Bash commands, Python and file edits inside an Anthropic-hosted container with no internet access.",
      "isolation": "Anthropic-hosted Linux container per request (\"secure, sandboxed container\"); the docs state full isolation from the host and other containers but do not name the container runtime or hypervisor",
      "isolation_short": "Hosted Linux container",
      "isolation_class": "container",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "Java",
        "C#",
        "PHP",
        "Ruby",
        "REST API",
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Containers expire 30 days after creation; after about 5 minutes of inactivity a container is checkpointed and restored when its ID is sent again; each tool call has a maximum execution time and each REPL cell in programmatic tool calling has a 90-second wall-clock limit",
      "max_session_short": "30 days (container)",
      "persistence": "Files written to the container persist across requests when you pass the container ID back; with code_execution_20260120 or later the Python REPL state (variable bindings) also persists; generated files are returned through the Files API",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented; each container has 1 CPU, 5 GiB RAM and 5 GiB of workspace storage",
      "networking": "Internet access is completely disabled and no outbound connections are allowed; packages cannot be installed at runtime, only the pre-installed libraries are available",
      "pricing": "Billed by execution time, separate from tokens: 1,550 free container-hours per organisation per month, then $0.05 per hour per container, with a 5-minute minimum per execution; free of additional charge in requests that also use the web search or web fetch tools (web_search_20260209 / web_fetch_20260209 or later)",
      "free_tier": "1,550 free container-hours per organisation per month",
      "compliance": "Not eligible for zero data retention (container data retained up to 30 days)",
      "regions": "Available on the Claude API, the Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry (hosted on Anthropic); no region selection documented",
      "best_for": "Data analysis, file generation and multi-step scripting inside a Claude API conversation without running your own sandbox",
      "overview": [
        "The code execution tool is a server-side tool in the Claude API: add `{\"type\": \"code_execution_20250825\", \"name\": \"code_execution\"}` to a Messages request and Claude can run Bash commands, write and edit files and execute Python in a container that Anthropic hosts. The container is a Linux x86_64 environment with Python 3.11, 1 CPU, 5 GiB of RAM and 5 GiB of workspace storage, with a fixed set of pre-installed libraries (pandas, numpy, scipy, scikit-learn, matplotlib, seaborn, openpyxl, pypdf and others) and command-line tools such as ripgrep, sqlite and 7zip. Internet access is completely disabled. Files are passed in and out through the Files API using `container_upload` blocks and `$OUTPUT_DIR`.",
        "Each request gets a fresh container unless you pass back the `container.id` from an earlier response, in which case files (and, with `code_execution_20260120` or later, Python REPL state) persist. Containers are checkpointed after about five minutes of inactivity and expire 30 days after creation. Three tool versions are current: `code_execution_20250825` (Bash and file operations), `code_execution_20260120` (adds REPL state persistence and programmatic tool calling) and `code_execution_20260521` (same runtime, with the 90-second per-cell limit described to the model); the legacy Python-only `code_execution_20250522` is documented for migration. The feature is marked stable and no beta header is required. Pricing is by execution time: 1,550 free container-hours per organisation per month, then $0.05 per container-hour with a five-minute minimum, and no charge when the request also uses the web search or web fetch tools. This record covers the hosted API tool only; Claude Code's local sandbox is a separate product."
      ],
      "notable": [
        "Marked stable in the Claude API docs; the three current tool versions are dated 2025-08-25, 2026-01-20 and 2026-05-21, and none requires a beta header",
        "Each organisation receives 1,550 free container-hours per month; usage beyond that is $0.05 per container-hour",
        "Images, video and audio that Claude produces in the container carry C2PA Content Credentials when downloaded through the Files API"
      ],
      "pros": [
        "You want Claude to analyse uploaded files or produce charts and documents without operating a sandbox yourself",
        "You need the execution environment to be fully offline for data-handling reasons",
        "You are already on the Claude API (or its AWS, Google Cloud or Microsoft Foundry distributions) and want a one-line opt-in",
        "You want REPL state and files to survive across turns by reusing a container ID"
      ],
      "cons": [
        "No internet access at all, so no package installation, git clones or API calls from inside the container",
        "Fixed resources: 1 CPU, 5 GiB RAM, 5 GiB disk; GPU and custom images are not documented",
        "Only usable through Claude; the container is not addressable as a general-purpose sandbox from your own code",
        "Container data is retained for up to 30 days, so the tool is not eligible for zero data retention"
      ],
      "verdict": "A server-side tool that gives Claude an Anthropic-hosted container for analysis and file work, suited to tasks that need neither network access nor custom environments.",
      "peers": [
        "openai-code-interpreter",
        "gemini-code-execution",
        "mistral-code-interpreter",
        "aws-agentcore",
        "azure-dynamic-sessions"
      ],
      "sources": [
        "https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "gemini-code-execution",
      "name": "Gemini API code execution",
      "company": "Google",
      "hq": {
        "city": "Mountain View",
        "region": "California",
        "country": "US",
        "source": "https://policies.google.com/terms"
      },
      "kind": "hyperscaler",
      "tier": 2,
      "alternatives": true,
      "website": "https://aistudio.google.com/docs/code-execution",
      "docs": "https://aistudio.google.com/docs/code-execution",
      "pricing_url": "https://aistudio.google.com/docs/code-execution",
      "github": null,
      "tagline": "Built-in Gemini API tool that lets the model write and run Python in a Google-hosted environment at no charge beyond tokens.",
      "isolation": "Not documented; Google describes a hosted Python environment with a fixed library set, a 30-second runtime limit and no package installation, but does not name the sandboxing technology",
      "isolation_short": "Hosted Python runtime",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "Java",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Each code execution has a maximum runtime of 30 seconds",
      "max_session_short": "30 s per execution",
      "persistence": "None documented; executions run inside a model turn, with multi-turn use via `previous_interaction_id`, and output files are returned as inline data",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Not documented; the docs say you cannot install your own libraries",
      "pricing": "No additional charge for enabling code execution; billed at the model's standard token rates (the prompt as input, generated code, execution results and final summary as output, intermediate tokens as input)",
      "free_tier": "No additional charge beyond model tokens",
      "compliance": null,
      "regions": null,
      "best_for": "Calculation, data analysis and chart generation inside a Gemini API call with no sandbox to operate",
      "overview": [
        "Code execution is a built-in tool of the Gemini API (the documentation now lives under AI Studio at aistudio.google.com/docs/code-execution). Passing `tools=[{\"type\": \"code_execution\"}]` lets the model generate Python, run it in a Google-hosted environment and iterate on the result; if the code errors the model may regenerate it up to five times. The environment ships a fixed library set including numpy, pandas, matplotlib, scipy, scikit-learn, sympy, opencv-python, pillow, tensorflow and seaborn, and you cannot install your own. Python is the only language that executes, graphs render with matplotlib only, and the maximum runtime of the environment is 30 seconds.",
        "Input files can be supplied inline or through the Files API (the docs say it works best with text and CSV files) and output files are returned as inline data. Multi-turn use goes through `previous_interaction_id`; the docs do not say whether interpreter state persists between turns, and do not mention internet access or the isolation technology. Code execution costs nothing beyond the model's standard token rates and can be combined with Grounding with Google Search and, on Gemini 3 and Gemini 4 models, with custom function calling. This is the Gemini API tool; Google's separate Agent Platform code execution service is covered by the google-code-execution record."
      ],
      "notable": [
        "No additional charge: the docs state \"There's no additional charge for enabling code execution\"",
        "Hard limits documented: Python only, 30-second maximum runtime, fixed library list, up to five automatic retries on error",
        "File input and graph output are documented as available since Gemini 2.0 Flash; image zoom, crop and annotation via code execution is documented for Gemini 3 Flash with Thinking enabled"
      ],
      "pros": [
        "You want Gemini to check its own maths or produce a chart without paying for or operating a sandbox",
        "You need a zero-setup option inside an existing Gemini API integration",
        "You are processing small CSV or text inputs that fit in the model's context window"
      ],
      "cons": [
        "30-second maximum runtime, and only Python code is executed",
        "No custom packages, no documented filesystem persistence and no documented network access",
        "Isolation technology is not documented",
        "The docs say that in some cases enabling code execution can lead to regressions in other areas of model output (for example, writing a story)"
      ],
      "verdict": "A built-in Python code execution tool for Gemini API calls at no charge beyond tokens, suited to short calculations and analysis within the documented 30-second runtime limit.",
      "peers": [
        "google-code-execution",
        "openai-code-interpreter",
        "anthropic-code-execution",
        "mistral-code-interpreter",
        "azure-dynamic-sessions"
      ],
      "sources": [
        "https://aistudio.google.com/docs/code-execution",
        "https://ai.google.dev/gemini-api/docs/code-execution"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "mistral-code-interpreter",
      "name": "Mistral code interpreter",
      "company": "Mistral AI",
      "hq": {
        "city": "Paris",
        "country": "FR",
        "source": "https://mistral.ai/legal"
      },
      "kind": "hyperscaler",
      "tier": 2,
      "alternatives": true,
      "website": "https://docs.mistral.ai/agents/tools/built-in/code_interpreter/",
      "docs": "https://docs.mistral.ai/agents/tools/built-in/code_interpreter/",
      "pricing_url": "https://mistral.ai/pricing",
      "github": null,
      "tagline": "Built-in tool of the Mistral Agents API that runs Python the agent writes in an isolated container.",
      "isolation": "Isolated container (\"safely execute code in an isolated container\"); the runtime and hypervisor are not named",
      "isolation_short": "Isolated container",
      "isolation_class": "container",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": null,
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": null,
      "pricing": "Not published: the Mistral pricing page says tool APIs are priced per call but lists no rate for code execution",
      "free_tier": null,
      "compliance": null,
      "regions": null,
      "best_for": "Letting a Mistral agent run short Python for maths, data analysis and plots without an external sandbox",
      "overview": [
        "Code interpreter is one of the built-in tools (connectors) of Mistral's Agents API, alongside web search, image generation and document library. You add `{\"type\": \"code_interpreter\"}` to the tools list when creating an agent, then drive it through the Conversations API; the model decides when to run code and the response carries a `tool.execution` entry with the code and its output. Mistral describes the environment as an isolated container for safely executing code, and the launch announcement (27 May 2025) calls it a secure sandboxed environment for Python; no further detail on the runtime, libraries, resource limits, networking or state persistence is published.",
        "The tool is not available on the Chat Completions endpoint, only on the Agents and Conversations APIs, which the SDKs expose under a `beta` namespace. Mistral's pricing page states that tool APIs are priced per call but does not list a figure for code execution. It is a code interpreter invoked by the model rather than a general-purpose sandbox API: no SDK for creating or addressing the container directly is documented. Mistral's agent tools page now carries the notice \"The Agents API v1 is about to be deprecated. Build new agents with the Agents SDK.\"; it gives no date."
      ],
      "notable": [
        "Launched with the Agents API on 27 May 2025 together with web search, image generation, document library and MCP tools",
        "Mistral's agent tools page carries the notice \"The Agents API v1 is about to be deprecated. Build new agents with the Agents SDK.\"; it gives no date",
        "Exposed under the `client.beta.agents` and `client.beta.conversations` SDK namespaces; not supported in Chat Completions",
        "Tool pricing is stated as per call on the pricing page, with no published rate for code execution"
      ],
      "pros": [
        "You are building on the Mistral Agents API and want code execution with one line of configuration",
        "You need an EU-headquartered model provider's native code tool rather than a third-party sandbox",
        "You want the model to decide when to run code and to see the code and its output in the conversation record"
      ],
      "cons": [
        "Runtime details (libraries, time and memory limits, networking, persistence) are not documented",
        "Price for the tool is not published",
        "Only reachable through the Agents and Conversations APIs; no API for creating or controlling the container directly is documented"
      ],
      "verdict": "A built-in code interpreter for agents on the Mistral Agents and Conversations APIs; its libraries, limits, networking and persistence are not documented.",
      "peers": [
        "openai-code-interpreter",
        "anthropic-code-execution",
        "gemini-code-execution",
        "azure-dynamic-sessions",
        "e2b"
      ],
      "sources": [
        "https://docs.mistral.ai/agents/tools/built-in/code_interpreter/",
        "https://mistral.ai/pricing",
        "https://mistral.ai/news/agents-api",
        "https://docs.mistral.ai/studio/agents/agent-tools/code_interpreter",
        "https://docs.mistral.ai/studio/agents/agent-tools",
        "https://docs.mistral.ai/inference/pricing"
      ],
      "verified_on": "2026-11-15"
    },
    {
      "id": "alibaba-agentbay",
      "name": "AgentBay",
      "company": "Alibaba Cloud",
      "hq": {
        "city": "Hangzhou",
        "region": "Zhejiang",
        "country": "CN",
        "entity": "Alibaba Group",
        "source": "https://www.alibabagroup.com/en-US/faqs-corporate-information"
      },
      "kind": "hyperscaler",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.alibabacloud.com/en/product/agentbay",
      "docs": "https://www.alibabacloud.com/help/en/agentbay/",
      "pricing_url": "https://www.alibabacloud.com/help/en/agentbay/product-overview/agentbay-billing-instructions",
      "github": "https://github.com/aliyun/wuying-agentbay-sdk",
      "tagline": "Alibaba Cloud's on-demand cloud sandboxes for agents, offering browser, Windows/Linux desktop, Android cloud-phone and code-space environments through one SDK.",
      "isolation": "Cloud-hosted environment instances described as sandboxed with \"safe isolation\"; the virtualisation technology is not named",
      "isolation_short": "Cloud environment instance",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "Java",
        "CLI",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Lifecycle policy per session: idle release timeout default 5 minutes (backend minimum 3), maximum runtime default 30 minutes from creation, or manual release only; `keep_alive()` resets the idle timer",
      "max_session_short": "30 min default; configurable",
      "persistence": "Sessions are temporary and lose data when released; Contexts (AgentBay-managed OSS directories) persist across sessions via Context Sync at start/end or Context Mount (beta, `aio-ubuntu-2404` image only); pause/resume is currently in whitelist-only access",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Basic bandwidth of 10 Mbps on all plans; premium bandwidth on Pro and Ultra billed at $0.123 per GB inbound",
      "pricing": "Plans plus pay-as-you-go: Basic free, Pro $149/month, Ultra $225/month (+$1.5 per extra concurrent session per month); CPU $0.027 per core-hour, memory $0.0113 per GB-hour billed per second, storage overage $0.000045 per GiB-hour; built-in agents metered in Credits at $0.0267 each",
      "free_tier": "Basic plan: free, 10 concurrent sessions, 100 GiB storage, system images only; one-time trial voucher of 100 points for compute",
      "compliance": null,
      "regions": "Deployed in Singapore (international site); a separate China site exists",
      "best_for": "Browser, desktop and mobile automation agents that need Chinese-cloud or Singapore-hosted environments",
      "overview": [
        "AgentBay (SDK package `wuying-agentbay-sdk`) is Alibaba Cloud's managed sandbox service for agents. A session is created from an image such as `browser_latest`, `code_latest`, `linux_latest` or a custom image, and the SDK exposes modules for command execution, file operations, Git, browser automation (Playwright and Puppeteer compatible), Computer Use on Windows and Linux desktops, and Mobile Use on cloud Android devices. Access is through Python, TypeScript, Go and Java SDKs, a CLI and an MCP server. Alibaba describes the environments as sandboxed with safe isolation but does not document the virtualisation layer.",
        "Sessions are explicitly temporary: data is lost when a session is released, and a lifecycle policy sets an idle timeout (default 5 minutes) and a maximum runtime (default 30 minutes) unless manual release is chosen. Durable state goes into Contexts, OSS-backed directories synced into a session path at start and end, or mounted live in a beta mode. Billing combines a plan (free Basic with 10 concurrent sessions, Pro at $149/month and Ultra at $225/month with 200 sessions each) with per-second compute at $0.027 per core-hour and $0.0113 per GB-hour; the trial page gives rough burn rates of about 0.71 points/hour for browser or code sessions and 1.43 for desktop or mobile. The international deployment is in Singapore."
      ],
      "notable": [
        "Pro plan $149/month and Ultra $225/month, each with 200 concurrent sessions; Basic is free with 10",
        "Default session lifecycle: 5-minute idle release, 30-minute maximum runtime, both configurable per session",
        "SDKs in four languages (Java artifact at 0.20.0) under Apache-2.0; a research paper is published as arXiv 2512.04367"
      ],
      "pros": [
        "You need browser, Windows/Linux desktop and Android environments from a single provider and SDK",
        "You run on Alibaba Cloud or need a Singapore-hosted sandbox",
        "You want a free plan with 10 concurrent sessions to prototype on",
        "You want Playwright or Puppeteer scripts to drive the hosted browser unchanged"
      ],
      "cons": [
        "Isolation technology is not documented",
        "Sessions default to a 30-minute maximum and lose all data on release unless Contexts are configured",
        "Compliance certifications and GPU support are not documented; pause/resume and Context Mount are beta or whitelist-only",
        "Requires an Alibaba Cloud account, and the free trial is not available to RAM users"
      ],
      "verdict": "A managed Alibaba Cloud sandbox service offering browser, Windows/Linux desktop, Android and code environments with published per-second compute pricing; the isolation technology is not documented.",
      "peers": [
        "aws-agentcore",
        "browserbase",
        "scrapybara",
        "e2b",
        "opensandbox",
        "cua"
      ],
      "sources": [
        "https://www.alibabacloud.com/en/product/agentbay",
        "https://www.alibabacloud.com/help/en/agentbay/",
        "https://www.alibabacloud.com/help/en/agentbay/product-overview/agentbay-billing-instructions",
        "https://www.alibabacloud.com/help/en/agentbay/product-overview/new-user-trial-instructions",
        "https://github.com/aliyun/wuying-agentbay-sdk",
        "https://github.com/aliyun/wuying-agentbay-sdk/blob/main/docs/guides/common-features/basics/session-management.md",
        "https://github.com/aliyun/wuying-agentbay-sdk/blob/main/docs/guides/common-features/basics/data-persistence.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "aio-sandbox",
      "name": "AIO Sandbox",
      "company": "Agent Infra (agent-infra GitHub organisation)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://sandbox.agent-infra.com",
      "docs": "https://sandbox.agent-infra.com",
      "pricing_url": null,
      "github": "https://github.com/agent-infra/sandbox",
      "tagline": "All-in-one agent sandbox image bundling a VNC browser, shell, filesystem, VS Code, Jupyter and an MCP server in a single Docker container.",
      "isolation": "Single Docker container per sandbox with a shared filesystem across its services; the quick start runs it with `--security-opt seccomp=unconfined`",
      "isolation_short": "Docker container",
      "isolation_class": "container",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "REST API",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "State lives in the container for as long as it runs; no snapshot or pause feature is documented",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Container listens on port 8080 (bound to 127.0.0.1 in the examples); port forwarding and preview for apps running inside; README advises keeping 8080 private behind a reverse proxy or Ingress when deployed in the cloud",
      "pricing": "Free to self-host; a hosted deployment through Volcengine Function Service (veFaaS) is billed by that platform",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": "Self-hosted anywhere; the veFaaS one-click deployment is limited to North China 2 (Beijing)",
      "best_for": "Agents that need a browser, terminal, editor and notebook in one container with a shared filesystem",
      "overview": [
        "AIO Sandbox is an open-source Docker image from the Agent Infra team (also offered as a one-click application on Volcengine Function Service) that packs everything an agent might need into one container: a Chromium browser with VNC and CDP access, a WebSocket shell, file APIs, VS Code Server, Jupyter, and an MCP server exposing browser, file, shell and markitdown tools. All components share one filesystem, so a file the shell writes is visible in the editor, the notebook and the browser. You run it with a single `docker run` (Compose and Kubernetes manifests are provided) and talk to it through a REST API at `/v1/...`, Python, TypeScript or Go SDKs, or MCP at `/mcp`.",
        "Isolation is the Docker container itself; the quick start disables the seccomp profile and the README recommends setting `SANDBOX_API_KEY`, otherwise the API, Jupyter and VNC are open. No snapshot, pause or multi-tenant control plane is documented: it is an environment image, and scheduling is left to you or to a hosting platform. A hosted path exists on Volcengine veFaaS (Beijing region only). The repository is Apache-2.0 with a pinned release line (1.11.0 cited in the README) and a companion paper, arXiv 2509.02544."
      ],
      "notable": [
        "Apache-2.0; roughly 6.3k GitHub stars and a Go SDK repository alongside Python and TypeScript packages",
        "Volcengine Function Service lists an \"All-in-One Sandbox\" one-click application restricted to the North China 2 (Beijing) region",
        "Accompanied by a research paper, arXiv 2509.02544"
      ],
      "pros": [
        "You want browser, shell, editor and notebook in one image without wiring services together",
        "You run computer-use or browser agents that also need to run code against the same files",
        "You need MCP tools for the sandbox out of the box",
        "You are on Volcengine or Alibaba Cloud Model Studio and want a hosted version of the same image"
      ],
      "cons": [
        "Isolation is the Docker container, and the documented run command sets seccomp to unconfined",
        "Snapshots, pause/resume and built-in orchestration are not documented; one container is one sandbox",
        "Services are open unless an API key is configured",
        "The repository identifies the maintainer as the \"Agent Infra team\"; it does not name a company"
      ],
      "verdict": "An all-in-one environment image for agents that combines browser, shell, files, VS Code, Jupyter and MCP in a single Docker container, for teams that run and schedule the containers themselves.",
      "peers": [
        "opensandbox",
        "docker-sandboxes",
        "openhands-runtime",
        "cua",
        "e2b"
      ],
      "sources": [
        "https://github.com/agent-infra/sandbox",
        "https://sandbox.agent-infra.com/",
        "https://github.com/agent-infra",
        "https://dev.to/bytedanceoss/introducing-aio-sandbox-all-in-one-sandbox-environment-for-ai-agents-18k0",
        "https://docs.volcengine.com/docs/6662/1851199",
        "https://help.aliyun.com/en/model-studio/sandbox-best-practice-aio"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "cube-sandbox",
      "name": "CubeSandbox",
      "company": "Tencent Cloud",
      "hq": {
        "entity": "Tencent",
        "city": "Shenzhen",
        "region": "Guangdong",
        "country": "CN",
        "source": "https://www.tencent.com/en-us/about.html"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://cubesandbox.com",
      "docs": "https://cubesandbox.com/guide/quickstart",
      "pricing_url": null,
      "github": "https://github.com/TencentCloud/CubeSandbox",
      "tagline": "Tencent Cloud's open-source, E2B-compatible sandbox platform that runs each sandbox in a RustVMM/KVM microVM with copy-on-write snapshots.",
      "isolation": "Dedicated microVM per sandbox with its own guest kernel, built on RustVMM and KVM (CubeHypervisor), with eBPF-based network isolation and egress filtering",
      "isolation_short": "RustVMM/KVM microVM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "REST API",
        "CLI"
      ],
      "cold_start": "README: sandboxes created in under 60 ms at single concurrency, average 67 ms / P95 90 ms / P99 137 ms at 50 concurrent creations, measured on bare metal; memory overhead claimed under 5 MB per sandbox",
      "cold_start_short": "<60 ms (claimed)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "CubeCoW copy-on-write snapshot engine with cloning, rollback and forking from saved state; AutoPause/AutoResume for idle sandboxes; cross-node pause/resume with an S3 backend (v0.7.0); creating sandboxes from snapshots is a preview; pluggable volumes (S3, host mounts, shared block)",
      "snapshots": true,
      "pause_resume": true,
      "gpu": false,
      "gpu_detail": "No — GPU sandboxes are on the roadmap, not available",
      "networking": "eBPF virtual switch (CubeVS) for inter-sandbox isolation and policy; CubeEgress gateway for L7 domain filtering, credential injection and auditing; per-sandbox traffic tokens",
      "pricing": "Free to self-host (Apache-2.0); no hosted offering is described",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": "Self-hosted; Terraform one-click deployment on Tencent Cloud",
      "best_for": "Self-hosted, E2B-compatible microVM sandboxes at high density on your own Linux hosts",
      "overview": [
        "CubeSandbox is a sandbox platform open-sourced by Tencent Cloud in April 2026. Each sandbox is a microVM with its own kernel managed by a Rust hypervisor (CubeHypervisor, built on RustVMM and KVM, with Cloud Hypervisor and Kata Containers acknowledged upstream), and a REST gateway (CubeAPI) and reverse proxy (CubeProxy) implement the E2B API so that the E2B Python, JavaScript or Go SDKs work by changing `E2B_API_URL`. A single-node install is a one-line script on x86_64 Linux (PVM mode needs no bare metal or nested virtualisation; ARM64 needs native KVM), and CubeMaster orchestrates multi-node clusters with Kubernetes and Terraform deployment options and a web console with a template store.",
        "The README claims sub-60 ms creation via snapshot clones that skip kernel boot, under 5 MB memory overhead and thousands of sandboxes per node. The CubeCoW engine provides copy-on-write snapshots, rollback and forking, and v0.7.0 (August 2026) added cross-node pause/resume backed by S3. Networking uses an eBPF virtual switch and an OpenResty egress gateway that filters domains and injects credentials so secrets never enter the sandbox. Minimum host is 4 cores, 8 GB RAM and 50 GB of XFS disk; GPU sandboxes and sandbox fault recovery are roadmap items. No hosted service is documented: cubesandbox.com covers self-build and cluster deployment only."
      ],
      "notable": [
        "Initial open-source release v0.1.0 on 20 April 2026; v0.7.0 on 28 August 2026 added cross-node pause/resume with an S3 backend",
        "Apache-2.0, listed in the CNCF Landscape, roughly 13k GitHub stars",
        "README claims 60 ms cold start at single concurrency and P99 137 ms at 50 concurrent creations on bare metal"
      ],
      "pros": [
        "You want E2B-compatible SDKs against infrastructure you run yourself",
        "You need microVM isolation with snapshot forking and rollback for agent branching or evaluation",
        "You run on Tencent Cloud or want a Terraform path onto it",
        "You need credential injection at the egress gateway so keys never reach the sandbox"
      ],
      "cons": [
        "Self-hosted; no managed service is documented, so you operate the hypervisor fleet",
        "Linux hosts only, with x86_64 PVM for the easy path and OpenCloudOS or TencentOS recommended",
        "No GPU sandboxes yet; snapshot-based creation and Kubernetes deployment are labelled preview in places",
        "First released in April 2026; the roadmap lists closing remaining gaps with the E2B specification"
      ],
      "verdict": "An open-source, E2B-compatible microVM sandbox platform for teams that want to run sandboxes on their own Linux hosts.",
      "peers": [
        "e2b",
        "opensandbox",
        "perplexity-space",
        "microsandbox",
        "firecracker",
        "kata-containers"
      ],
      "sources": [
        "https://github.com/TencentCloud/CubeSandbox",
        "https://cubesandbox.com",
        "https://cubesandbox.com/guide/quickstart",
        "https://github.com/TencentCloud/CubeSandbox/releases"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "opensandbox",
      "name": "OpenSandbox",
      "company": "Agentic AI Foundation (contributed by Alibaba)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://open-sandbox.ai",
      "docs": "https://open-sandbox.ai",
      "pricing_url": null,
      "github": "https://github.com/opensandbox-group/OpenSandbox",
      "tagline": "Apache-2.0 sandbox runtime for agents that starts on Docker, scales on Kubernetes and offers a Firecracker \"Fast Sandbox\" with pause/resume.",
      "isolation": "Docker containers locally and Kubernetes pods in clusters, plus a Firecracker microVM \"Fast Sandbox\" runtime on Kubernetes",
      "isolation_short": "Containers or Firecracker",
      "isolation_class": "mixed",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Java",
        "Kotlin",
        "Go",
        "C#",
        "CLI",
        "MCP server",
        "REST API"
      ],
      "cold_start": "Website: Firecracker warm creation 64 ms P50 (serial) and 125.4 ms P99 (10 concurrent) on the Fast Sandbox runtime",
      "cold_start_short": "64 ms P50 (claimed)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Fast Sandbox (Firecracker) sandboxes can be paused and resumed with memory and disk state preserved; a snapshot API is documented for Docker and compatible BatchSandbox workloads, and Fast Sandbox uses a template catalogue",
      "snapshots": null,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "GPU resource limits are documented for the Docker and Kubernetes runtimes; the secure container guide lists no GPU support for Kata with Firecracker",
      "networking": "Unified ingress gateway for inbound traffic; per-sandbox outbound egress policies; Credential Vault lets agents call external services without real credentials entering the sandbox",
      "pricing": "Free to self-host (Apache-2.0); no managed service listed",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": "Self-hosted; images published to Docker Hub, GHCR and Alibaba Cloud Container Registry",
      "best_for": "Teams that want an open, multi-language sandbox API they can run from a laptop Docker daemon up to a Kubernetes cluster",
      "overview": [
        "OpenSandbox is an open-source sandbox runtime for agents, now governed as an Agentic AI Foundation project under LF Projects; Alibaba's origin shows in the SDK namespaces (`com.alibaba.opensandbox`, `@alibaba-group/opensandbox`, `github.com/alibaba/OpenSandbox`) and an Alibaba Cloud registry mirror. A server (`uvx opensandbox-server`) exposes a unified lifecycle API for creating sandboxes, running commands, managing files and executing code, with SDKs for Python, TypeScript, Java/Kotlin, Go and C#, an `osb` CLI and an MCP server for clients such as Claude Code and Cursor. Docker is the local runtime; Kubernetes adds resource pools and batch creation; and a Fast Sandbox runtime on Kubernetes uses Firecracker microVMs, with pause/resume that preserves memory and disk.",
        "The project publishes integration examples for coding agents (Claude Code, Codex CLI, DeerFlow), a code interpreter, Playwright and Chrome, desktop environments and Harbor evaluations. Security features include an ingress gateway, per-sandbox egress policy and a Credential Vault. The website quotes 64 ms P50 warm creation on Firecracker. Images are published to Docker Hub, GHCR and Alibaba Cloud's registry, though the README warns that current umbrella releases lack the Cosign signing and provenance that earlier per-component releases had and advises pinning by digest. The architecture docs describe a snapshot API for Docker and compatible BatchSandbox workloads and GPU resource limits; the secure container guide lists no GPU support for Kata with Firecracker."
      ],
      "notable": [
        "Contributed by Alibaba to the Agentic AI Foundation alongside Agent Skills, Open Responses and OpenEnv",
        "Apache-2.0 with roughly 17k GitHub stars and SDKs in five languages plus CLI and MCP",
        "Website claims 64 ms P50 and 125.4 ms P99 (10 concurrent) warm creation on the Firecracker Fast Sandbox runtime"
      ],
      "pros": [
        "You want one API and SDK set that works on Docker locally and Kubernetes in production",
        "You need Firecracker-level isolation with pause/resume on Kubernetes",
        "You run polyglot teams (Java, C#, Go, TypeScript, Python) that all need the same sandbox client",
        "You want foundation governance rather than a single-vendor project"
      ],
      "cons": [
        "Default runc runtime is container isolation; gVisor and Kata secure runtimes need extra configuration, and Kata with Firecracker is not supported in Docker mode",
        "Public snapshot support does not extend to every Kubernetes provider or VM runtime, and GPU is not supported with Kata on Firecracker",
        "Umbrella releases currently lack signing and provenance evidence (tracked issue), so images should be pinned by digest",
        "No managed offering is documented; you run the control plane yourself"
      ],
      "verdict": "An open sandbox API with SDKs in five languages that runs on Docker locally and Kubernetes in clusters, with Firecracker microVM isolation available on Kubernetes, for teams that self-host.",
      "peers": [
        "e2b",
        "cube-sandbox",
        "aio-sandbox",
        "gke-agent-sandbox",
        "daytona",
        "microsandbox"
      ],
      "sources": [
        "https://github.com/opensandbox-group/OpenSandbox",
        "https://open-sandbox.ai/",
        "https://github.com/alibaba/OpenSandbox",
        "https://www.linuxfoundation.org/press/agentic-ai-foundation-announces-four-new-projects-agent-skills-open-responses-openenv-and-opensandbox",
        "https://github.com/opensandbox-group/OpenSandbox/blob/main/docs/community/release-verification.md",
        "https://github.com/opensandbox-group/OpenSandbox/issues/2463",
        "https://open-sandbox.ai/architecture/",
        "https://open-sandbox.ai/guides/secure-container"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "perplexity-space",
      "name": "SPACE",
      "company": "Perplexity",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.perplexity.ai/hub/legal/terms-of-service"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://space.docs.perplexity.ai",
      "docs": "https://space.docs.perplexity.ai",
      "pricing_url": null,
      "github": "https://github.com/perplexityai/SPACE",
      "tagline": "Perplexity's open-source sandbox platform that runs code in Firecracker microVMs with pause, suspend-to-snapshot and an E2B compatibility layer.",
      "isolation": "Firecracker microVM per sandbox with its own writable filesystem (\"Firecracker isolates guest execution\")",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Python",
        "REST API",
        "CLI"
      ],
      "cold_start": "No boot time figure published; warm pools keep booted sandboxes ready and the native create readiness budget is 5 seconds",
      "cold_start_short": null,
      "max_session": "Default idle suspend after 6 hours without access or command activity; lifecycle operations have a 30-minute deadline",
      "max_session_short": "6 h idle suspend",
      "persistence": "Paused keeps VM memory resident on the worker (not durable); suspended writes a snapshot of memory, runtime state and root filesystem and releases VM resources; wake restores either; snapshots, checkpoint and rollback are incubating and hidden unless enabled per organisation",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Egress governed by NetworkPolicyV1 destination rules with optional per-policy TLS interception; operator egress profiles; in the Kind install sandbox pods are restricted to public IPv4 and cluster DNS",
      "pricing": "Free to self-host (Apache-2.0); no hosted offering described",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": "Self-hosted; local Kind cluster install documented, with an AWS EKS reference deployment in the repository",
      "best_for": "Operators who want a Firecracker sandbox control plane with asynchronous lifecycle operations and E2B SDK compatibility",
      "overview": [
        "SPACE is a sandbox platform published by Perplexity under Apache-2.0. A template supplies the starting environment and a sandbox is an instance with its own Firecracker microVM and writable filesystem; applications drive it over an HTTP API with server-sent events, with command output flowing over a separate data plane and lifecycle changes returned as operation IDs to poll. An API server backed by PostgreSQL stores desired state and controllers reconcile it, either directly against worker services or through a Kubernetes adapter that creates sandbox pods. The E2B Compatibility Wrapper exposes an E2B-style API tested against the E2B Python SDK, with limitations listed in the repository.",
        "Sandboxes can be paused (memory stays resident, not durable), suspended (a snapshot of memory, runtime state and root filesystem is written and VM resources released) and woken; idle sandboxes suspend after 6 hours by default. Snapshot export, checkpoint and rollback are incubating features hidden unless enabled for an organisation, and file writes persist at lifecycle boundaries rather than per write. Egress is policy-driven with optional TLS interception. The documented install targets an apt-based x86_64 Linux host with KVM, Docker and Kind; the installation guide says its instructions do not run directly on macOS. The docs site covers version 0.8.2; no hosted version is documented."
      ],
      "notable": [
        "Apache-2.0; documentation covers SPACE version 0.8.2 and the repository has roughly 2.1k GitHub stars",
        "Default idle suspend is 6 hours; suspend produces a durable snapshot of memory, runtime state and root filesystem",
        "E2B Compatibility Wrapper is tested against the E2B Python SDK, with supported behaviour and gaps documented in docs/ecw.md"
      ],
      "pros": [
        "You want Firecracker isolation with pause, suspend and wake semantics spelled out",
        "You need egress policy with TLS interception and per-organisation feature gating",
        "You have existing E2B Python SDK code and want to point it at your own cluster",
        "You run on Kubernetes (an EKS reference exists) or can run the direct adapter without it"
      ],
      "cons": [
        "Linux x86_64 hosts with KVM only; the install does not run on macOS",
        "Boot time figures are not published and no hosted version is documented",
        "Snapshots, checkpoint and rollback are incubating and hidden by default",
        "Only a Python example and the E2B wrapper are documented as client paths; first-party SDKs are not documented"
      ],
      "verdict": "An open-source Firecracker sandbox control plane for operators who self-host on Linux with Kubernetes, with an E2B compatibility wrapper for existing E2B Python SDK code.",
      "peers": [
        "e2b",
        "cube-sandbox",
        "firecracker",
        "microsandbox",
        "opensandbox",
        "arrakis"
      ],
      "sources": [
        "https://github.com/perplexityai/SPACE",
        "https://space.docs.perplexity.ai/",
        "https://github.com/perplexityai/SPACE/blob/main/docs/installation.md",
        "https://github.com/perplexityai/SPACE/blob/main/docs/architecture.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "nvidia-openshell",
      "name": "OpenShell",
      "company": "NVIDIA",
      "hq": {
        "city": "Santa Clara",
        "region": "California",
        "country": "US",
        "source": "https://www.nvidia.com/en-us/contact/"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://docs.nvidia.com/openshell/latest/",
      "docs": "https://docs.nvidia.com/openshell/latest/",
      "pricing_url": null,
      "github": "https://github.com/NVIDIA/OpenShell",
      "tagline": "NVIDIA's open-source policy-enforcing runtime that runs each agent in a sandbox with Landlock, seccomp and a supervisor that mediates every outbound connection and credential.",
      "isolation": "Workload runs as a non-root identity with no Linux capabilities, confined by Landlock (filesystem) and seccomp (process and network interception), inside a Docker or Podman container with networking off, a Kubernetes pod with a NetworkPolicy, or a VM guest with no network device",
      "isolation_short": "Landlock/seccomp in container or VM",
      "isolation_class": "mixed",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "Rust",
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Filesystem and process policy are fixed at sandbox creation and changing them requires recreating the sandbox; sandboxes can be stopped and started with a persistent workspace retained; no snapshot or pause feature is documented",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Documented: `--gpu` requests GPU resources (NVIDIA CDI devices on Docker and Podman, an `nvidia.com/gpu` limit on Kubernetes, one GPU on the MicroVM driver)",
      "networking": "Deny-all outer fence; the sandbox intercepts TCP opens and DNS queries with seccomp and forwards them to the supervisor, which allows only policy-approved destinations and injects credentials only for approved endpoints; network policy is hot-reloadable",
      "pricing": "Free to self-host (Apache-2.0)",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": "Self-hosted: local installer on Linux and macOS (Apple Silicon), Windows via WSL 2 (experimental), Kubernetes via Helm, or a Brev Launchable in the cloud",
      "best_for": "Running coding agents such as Claude Code, Codex and OpenCode under an auditable, version-controlled policy for files, processes, network and credentials",
      "overview": [
        "OpenShell is NVIDIA's open-source runtime for autonomous agents; its README describes it as a sandbox that enforces a policy you declare per agent. A gateway (control plane) manages sandbox lifecycles and policy, a supervisor on the trusted side of the boundary makes every policy decision and holds credentials, and the sandbox itself runs the agent as a non-root identity with no Linux capabilities, confined by Landlock for files and seccomp for process and network operations. Every outbound TCP open and DNS query is intercepted and forwarded to the supervisor, which checks it against policy and adds credentials only for approved endpoints, so the agent never sees a real key. Runtimes are Docker or Podman containers with networking disabled, Kubernetes pods with NetworkPolicy, or a VM guest with no network device over vsock.",
        "Policy is declarative YAML covering filesystem, process and network rules; agent-proposed network changes pass through a formal policy prover and anything that adds risky reach is held for human review. You create sandboxes with the CLI (`openshell sandbox create --name demo`) or SDKs for Python, TypeScript, Go and Rust; the default image is minimal Ubuntu with no agent installed, and the docs name Claude Code, OpenCode, Codex and GitHub Copilot CLI as supported agents. The 0.1.x line (v0.1.7 at the time of verification) is described as having a stable release cadence. OpenShell qualifies as a sandbox because it provisions and confines the environment the agent runs in; its distinguishing feature is the policy and credential layer rather than a new isolation primitive, and it supports pluggable isolation backends including microVMs."
      ],
      "notable": [
        "Apache-2.0 with roughly 21k GitHub stars; documentation covers v0.1.8 and describes the 0.1.x line as a stable release cadence",
        "Agents never receive real credentials: the supervisor injects them only into requests bound for policy-approved endpoints",
        "Proposed policy changes are checked by a formal policy prover (`openshell-prover`) before auto-approval"
      ],
      "pros": [
        "You want coding agents to run with least-privilege file, process and network access that is enforced in the kernel",
        "You need credentials kept out of the agent's reach while it still calls approved APIs",
        "You run on Linux or Apple Silicon locally and on Kubernetes in production, with the same policy model",
        "You want policy as version-controlled YAML for audit and compliance"
      ],
      "cons": [
        "Auto-detected runtimes are containers (Kubernetes, Podman, Docker); the MicroVM driver must be enabled explicitly",
        "Snapshots and pause/resume are not documented; filesystem and process policy are locked at sandbox creation",
        "Windows support is experimental via WSL 2"
      ],
      "verdict": "An open-source runtime that runs agents under declarative file, process, network and credential policy, for teams whose priority is controlling what an agent can reach.",
      "peers": [
        "anthropic-srt",
        "docker-sandboxes",
        "nono",
        "bubblewrap",
        "opensandbox",
        "gemini-cli-sandbox"
      ],
      "sources": [
        "https://github.com/NVIDIA/OpenShell",
        "https://docs.nvidia.com/openshell/latest/index.html",
        "https://docs.nvidia.com/openshell/latest/about/architecture",
        "https://docs.nvidia.com/openshell/latest/about/how-it-works",
        "https://docs.nvidia.com/openshell/latest/extensibility/isolation-backends",
        "https://docs.nvidia.com/openshell/latest/how-it-works/sandboxes/runtimes",
        "https://docs.nvidia.com/openshell/how-it-works/sandboxes/overview"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "windows-sandbox",
      "name": "Windows Sandbox",
      "company": "Microsoft",
      "hq": {
        "city": "Redmond",
        "region": "Washington",
        "country": "US",
        "source": "https://www.microsoft.com/en-us/about/office-locations"
      },
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://learn.microsoft.com/en-us/windows/security/application-security/application-isolation/windows-sandbox/",
      "docs": "https://learn.microsoft.com/en-us/windows/security/application-security/application-isolation/windows-sandbox/",
      "pricing_url": null,
      "github": null,
      "tagline": "Disposable Windows desktop built into Windows Pro, Enterprise and Education, running on the Microsoft hypervisor with a separate kernel and wiped on close.",
      "isolation": "Hyper-V based lightweight virtual machine with its own kernel, isolated from the host by the Microsoft hypervisor; optional Protected Client mode adds AppContainer isolation to the RDP session",
      "isolation_short": "Hyper-V lightweight VM",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": "Docs: \"takes a few seconds to launch\"",
      "cold_start_short": "A few seconds",
      "max_session": "No limit documented; the sandbox lives until the window is closed, and since Windows 11 22H2 survives reboots initiated inside it",
      "max_session_short": "Until closed",
      "persistence": "None: closing the sandbox deletes all software, files and state and each launch is a pristine instance; only writes to mapped host folders with write permission persist",
      "snapshots": false,
      "pause_resume": null,
      "gpu": true,
      "gpu_detail": "Yes — virtualised GPU (vGPU) enabled by default on non-Arm64 devices; can be disabled in favour of WARP software rendering",
      "networking": "Enabled by default through the Hyper-V default switch and a virtual NIC; can be disabled per launch in a .wsb configuration file (the docs warn that networking can expose untrusted applications to the internal network)",
      "pricing": "Included with Windows Pro, Pro Education/SE, Enterprise (E3/E5) and Education (A3/A5); not available on Windows Home",
      "free_tier": "Included in the Windows licence",
      "compliance": null,
      "regions": null,
      "best_for": "Running an untrusted Windows application or a Windows-targeting coding agent on a disposable copy of the host OS",
      "overview": [
        "Windows Sandbox (WSB) is an optional Windows feature that launches a lightweight, disposable Windows desktop in a Hyper-V virtual machine with its own kernel. Everything needed ships with Windows Pro, Enterprise and Education (not Home); you enable the `Containers-DisposableClientVM` feature, click Windows Sandbox in the Start menu and get a pristine instance in a few seconds, with 4 GB of memory, vGPU, networking and clipboard redirection on by default. Closing the window deletes all software, files and state. Host-installed software is not visible inside, so an agent or tool has to be installed in the sandbox or mapped in from a host folder.",
        "Configuration is through XML `.wsb` files (Windows 10 build 18342 or Windows 11) that control vGPU, networking, mapped folders with read-only or write access, a logon command, audio and video input, printer and clipboard redirection, Protected Client (AppContainer) mode and memory in MB (minimum 2048 MB). Prerequisites are AMD64 or Arm64 (Windows 11 22H2+), virtualisation enabled in firmware, at least 4 GB RAM, 1 GB of disk and two CPU cores. Only one sandbox instance can run at a time; its value for agents is a free, OS-vendor-maintained throwaway Windows machine with a scripted startup command."
      ],
      "notable": [
        "Built into Windows 10 version 1903 and later and Windows 11 Pro, Enterprise and Education; not supported on Windows Home",
        "Since Windows 11 version 22H2 data persists through restarts initiated inside the sandbox; since 24H2 inbox store apps such as Notepad and Terminal are temporarily unavailable inside it",
        "Microsoft Learn page last updated 29 March 2026"
      ],
      "pros": [
        "You want a disposable Windows environment with no separate VM image to maintain",
        "You need GPU-accelerated graphics or a real Windows desktop for a computer-use agent",
        "You want a one-file `.wsb` configuration that maps a project folder in and runs a startup script",
        "You are on a Windows Pro or Enterprise fleet and need something IT can enable without new software"
      ],
      "cons": [
        "Only one instance can run at a time, and no API beyond the CLI and .wsb configuration files is documented",
        "Nothing persists between launches except writes to mapped folders",
        "Networking and clipboard sharing are on by default and must be disabled explicitly for hostile workloads",
        "Windows Home is not supported and the feature requires virtualisation enabled in firmware (nested virtualisation inside a VM)"
      ],
      "verdict": "A disposable Windows desktop included with Windows Pro, Enterprise and Education that runs one instance at a time, suited to single-agent or manual use.",
      "peers": [
        "apple-container",
        "docker-sandboxes",
        "anthropic-srt",
        "microsandbox",
        "kasm-workspaces"
      ],
      "sources": [
        "https://learn.microsoft.com/en-us/windows/security/application-security/application-isolation/windows-sandbox/",
        "https://learn.microsoft.com/en-us/windows/security/application-security/application-isolation/windows-sandbox/windows-sandbox-install",
        "https://learn.microsoft.com/en-us/windows/security/application-security/application-isolation/windows-sandbox/windows-sandbox-configure-using-wsb-file"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "apple-container",
      "name": "Apple container",
      "company": "Apple",
      "hq": {
        "city": "Cupertino",
        "region": "California",
        "country": "US",
        "source": "https://www.apple.com/contact/"
      },
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/apple/container",
      "docs": "https://github.com/apple/container/blob/main/docs/technical-overview.md",
      "pricing_url": null,
      "github": "https://github.com/apple/container",
      "tagline": "Apple's open-source CLI that runs each Linux container in its own lightweight virtual machine on Apple silicon Macs.",
      "isolation": "One lightweight Linux VM per container via Apple's Containerization Swift package and Virtualization.framework, with an optimised kernel, minimal root filesystem and a small `vminitd` init",
      "isolation_short": "Lightweight VM per container",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": false,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": "Containerization README: sub-second container start times attributed to an optimised kernel configuration and minimal root filesystem; `container` docs say boot times are comparable to containers in a shared VM, with no figures",
      "cold_start_short": "Sub-second (claimed)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Standard container lifecycle (start, stop, remove) on OCI images; no pause feature is documented; `container commit` creates an image from a container's filesystem",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Containers attach to a vmnet-managed virtual network and receive their own IP addresses (default subnet 192.168.64.1/24 on macOS 15); container-to-container communication and `container network` commands require macOS 26",
      "pricing": "Free (Apache-2.0); requires an Apple silicon Mac",
      "free_tier": "Free",
      "compliance": null,
      "regions": null,
      "best_for": "Running Linux containers for coding agents on an Apple silicon Mac with VM-grade isolation instead of a shared VM",
      "overview": [
        "Apple's `container` is an open-source, Swift-written command-line tool for creating and running Linux containers on Apple silicon Macs. Unlike Docker Desktop's single shared Linux VM, each container runs in its own lightweight virtual machine created through Apple's Containerization package on Virtualization.framework: the VM boots an optimised kernel and a minimal root filesystem with a small gRPC-over-vsock init (`vminitd`), and only the data a container needs is mounted into it, giving, in Apple's words, the isolation properties of a full VM with lower memory use. It consumes and produces standard OCI images, so images built elsewhere run and images built here push to ordinary registries; Rosetta 2 supports linux/amd64 images.",
        "Installation is a signed package from GitHub releases followed by `container system start`; a launch agent (`container-apiserver`) and XPC helpers manage images, the vmnet network and one `container-runtime-linux` process per container. macOS 26 is required for full function (macOS 15 runs with limitations such as no container-to-container networking) and Xcode 26 is needed to build Containerization. The project is under active development with product (not semantic) versions; release 1.7.1 was current at verification. No client SDK is documented beyond the CLI and the `container-apiserver` XPC API, and no pause feature is documented; the Containerization package also describes a Linux backend on cloud-hypervisor and KVM that shares the same guest contract."
      ],
      "notable": [
        "Apache-2.0; the `container` repository has roughly 51k GitHub stars and release 1.7.1 was current at verification; Containerization 0.1.0 is its first release with source-stability guarantees",
        "Requires an Apple silicon Mac and macOS 26 or later; the README says older macOS versions are not supported, and the technical overview does not recommend macOS 15",
        "Containerization README claims sub-second container start times from an optimised kernel and minimal root filesystem"
      ],
      "pros": [
        "You want per-container VM isolation on a Mac without running a shared Linux VM",
        "You need standard OCI images to work unchanged on Apple silicon, including amd64 via Rosetta",
        "You want each container to have its own IP so services are reachable without port forwarding",
        "You prefer an OS-vendor tool under Apache-2.0 to a commercial desktop container product"
      ],
      "cons": [
        "Requires a Mac with Apple silicon and macOS 26 or later; other hosts are not supported",
        "Pause/resume and a client SDK beyond the CLI and an XPC API are not documented",
        "Memory freed inside a container's VM is not returned to macOS (ballooning only partially supported), so long-lived heavy containers may need restarts",
        "The project notes the initial release lacks many common containerisation features and that breaking changes may occur"
      ],
      "verdict": "An Apple-maintained open-source CLI that runs each Linux container in its own lightweight VM, for developers on Apple silicon Macs with macOS 26 who want per-container VM isolation locally.",
      "peers": [
        "docker-sandboxes",
        "windows-sandbox",
        "microsandbox",
        "libkrun",
        "anthropic-srt",
        "boxlite"
      ],
      "sources": [
        "https://github.com/apple/container",
        "https://github.com/apple/container/blob/main/docs/technical-overview.md",
        "https://github.com/apple/container/releases",
        "https://github.com/apple/containerization",
        "https://github.com/apple/container/blob/main/docs/command-reference.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "digitalocean-harness-runtime",
      "name": "DigitalOcean Harness Runtime",
      "company": "DigitalOcean",
      "hq": {
        "city": "Broomfield",
        "region": "Colorado",
        "country": "US",
        "source": "https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001582961&type=10-K&dateb=&owner=include&count=10"
      },
      "kind": "hyperscaler",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.digitalocean.com/products/managed-agents",
      "docs": "https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/",
      "pricing_url": "https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/details/pricing/",
      "github": null,
      "tagline": "DigitalOcean's managed microVM sandboxes that run unmodified agents such as Claude Code, Codex or LangGraph, composed from a YAML environment of adapter, size, permissions, secrets, tools and egress rules.",
      "isolation": "One Firecracker microVM per session with its own kernel; DigitalOcean says isolation is enforced by the hypervisor rather than container namespaces",
      "isolation_short": "Firecracker microVM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "CLI",
        "REST API"
      ],
      "cold_start": "Creation to a first agent response in \"under a couple of seconds\"; paused sessions resume in about 300 ms according to the product page intro (its FAQ says about 200 ms) (vendor claims)",
      "cold_start_short": "<2 s (claim)",
      "max_session": "No fixed session length documented; sandboxes pause after an idle timeout (default 15 minutes, configurable, or disabled with `keep_warm: true`) and resume on the next input, exec or port-forward; each session has a monthly active-compute cap of 744 hours",
      "max_session_short": "No fixed limit; idle pause after 15 min",
      "persistence": "Pausing keeps memory, files and running processes with compute stopped; checkpoints capture the whole machine for fork and rollback; Persistent Workspaces (public preview, 1 to 100 GiB) save `/workspace` when a session is removed or checkpointed so a later session can attach it",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Outbound access is unrestricted by default; an `egress` hostname allowlist confines the sandbox to the listed hosts. Ports reach your machine only through `doctl harness-runtime port-forward`, and nothing is published to the internet. Bandwidth is 32 to 512 Mbps by size; private VPC connectivity and VPC peering are not supported",
      "pricing": "Billed per second of active use: CPU $0.044 per vCPU-hour on CPU actually consumed, memory $0.0095 per GB-hour allocated, so the default 2 vCPU / 4 GB sandbox is at most $0.126 per hour; session storage, checkpoints and custom templates $0.05 per GiB-month; public egress $0.01 per GiB; paused sessions incur no compute charges. Optional Pro ($50/month) and Team ($200/month) plans pre-buy usage at a 15% or 20% discount",
      "free_tier": "$5 promotional credit for new customers; Persistent Workspaces are free during public preview",
      "compliance": null,
      "regions": "RIC1 only, per the availability page",
      "best_for": "Teams that want to run unmodified coding agents such as Claude Code or Codex in managed microVMs with permission policies, checkpoints and forks, especially if they already use DigitalOcean",
      "overview": [
        "Harness Runtime is the sandbox half of DigitalOcean Managed Agents, which entered public preview for all users on 21 September 2026 alongside Action Gateway, a managed MCP endpoint in front of more than 16,000 tools. Each session gets its own Firecracker microVM with its own kernel. The agent runs inside it unmodified through an adapter: Codex, Claude Code and OpenCode for coding, and LangGraph, CrewAI, Hermes and OpenHands as general frameworks. Harness Runtime translates each agent's native output into one event stream.",
        "The product is composable. An environment is an immutable YAML spec, or a Control Panel form, that combines the adapter (`harness`), a size from 1 vCPU / 1 GB to 16 vCPU / 32 GB, Agent Skills, secrets, permission rules that allow, ask about or deny shell commands and file paths, MCP and Action Gateway tools, and an egress allowlist. Teams can register their own OCI image as a template, rebased onto one of DigitalOcean's base images. Since 26 October 2026, `harness: none` boots a bare Ubuntu 24.04 sandbox that you drive with `exec`, file upload and download, and port forwarding through `doctl`. Checkpoint, fork and rollback work as they do for other sessions.",
        "Sessions pause after 15 idle minutes by default, keeping memory, files and running processes, and resume on the next input. Checkpoints capture the whole machine, so you can fork a session or roll it back. Billing is per second of active use, with CPU charged on consumption rather than allocation, and pausing stops compute charges. Clients are `doctl`, the REST API, and PyDo, DoTs and Godo client libraries for Python, TypeScript and Go. A Cuttlefish iOS app lets you approve tool calls from a phone. The availability page lists a single datacenter, RIC1."
      ],
      "notable": [
        "One Firecracker microVM per session, with checkpoints, forks and rollback of the whole machine",
        "Runs Claude Code, Codex, OpenCode, LangGraph, CrewAI, Hermes and OpenHands unmodified through adapters, or a bare sandbox with `harness: none`",
        "CPU billed on consumption at $0.044 per vCPU-hour, so an agent waiting on a model costs little; up to 100 concurrent sessions per team"
      ],
      "pros": [
        "You want to run an off-the-shelf agent such as Claude Code or Codex in a managed microVM without wrapping it yourself",
        "You need allow, ask and deny permission policies, approvals and an egress allowlist declared per environment",
        "You want pause, checkpoint, fork and rollback for long-running agent sessions",
        "You already use DigitalOcean for databases, storage or inference"
      ],
      "cons": [
        "The whole product is in public preview, and Persistent Workspaces and Signals are preview features within it",
        "Only one datacenter (RIC1) is listed, with no VPC peering or private connectivity",
        "Compliance certifications and GPU support are not documented for Harness Runtime",
        "Egress is unrestricted unless you set an allowlist"
      ],
      "verdict": "A composable, agent-first microVM service with checkpoint and fork support and consumption-based CPU billing, currently in public preview and listed in one datacenter (RIC1).",
      "peers": [
        "aws-agentcore",
        "e2b",
        "daytona",
        "fly-sprites",
        "morph",
        "blaxel"
      ],
      "sources": [
        "https://www.digitalocean.com/products/managed-agents",
        "https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/",
        "https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/details/features/",
        "https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/concepts/architecture/",
        "https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/concepts/environments/",
        "https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/details/pricing/",
        "https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/details/limits/",
        "https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/details/availability/",
        "https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/how-to/run-sandbox/",
        "https://docs.digitalocean.com/release-notes/agent-harness-runtime/",
        "https://www.digitalocean.com/blog/microvms-for-agents"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "aws-nitro-sandboxes",
      "name": "AWS Nitro Sandboxes",
      "company": "Amazon Web Services",
      "hq": {
        "city": "Seattle",
        "region": "Washington",
        "country": "US",
        "source": "https://aws.amazon.com/privacy/"
      },
      "kind": "hyperscaler",
      "tier": 2,
      "alternatives": true,
      "website": "https://aws.amazon.com/blogs/compute/nitro-sandboxes-hardware-isolation-for-every-eks-pod-ecs-task-and-ec2-workload/",
      "docs": "https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/amazon-ec2-nitro-sandboxes.html",
      "pricing_url": null,
      "github": null,
      "tagline": "An Amazon EC2 feature that runs each EKS pod, ECS task or Firecracker workload in its own lightweight VM beside your instance, with the Nitro hypervisor enforcing the boundary.",
      "isolation": "Each sandbox is a lightweight VM with its own kernel that runs on the Nitro Hypervisor beside the parent EC2 instance and is launched through Firecracker; the Nitro System isolates sandboxes from each other, and in confidential mode it also removes sandbox memory from the parent instance",
      "isolation_short": "Nitro hypervisor VM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "Kubernetes RuntimeClass",
        "ECS task definition",
        "Firecracker API",
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "No session limit is documented: a sandbox runs until it powers off, its init process exits or its Firecracker process stops, and every sandbox stops when the parent instance stops, reboots or terminates",
      "max_session_short": "No limit; tied to parent instance",
      "persistence": "No snapshot and restore, pause and resume or live migration. State lives on block devices backed by instance store or EBS volumes of the parent instance, or on S3 and EFS over the network; on EKS, pods can mount EBS and EFS volumes through the CSI drivers",
      "snapshots": false,
      "pause_resume": false,
      "gpu": false,
      "gpu_detail": "No — GPU passthrough into a Nitro Sandbox is currently unavailable; on accelerated instances a sandbox runs CPU workloads only",
      "networking": "Each sandbox runs its own network stack. On EKS, pods keep a VPC IP through the VPC CNI, and an annotation boots a pod with no network device at all; on ECS each task gets an elastic network interface (`awsvpc` mode). With Firecracker you wire a TAP device yourself and must block the parent's instance metadata service (169.254.169.254), which a routed sandbox can otherwise reach",
      "pricing": "No additional charge: sandboxes run on the vCPUs and memory of the parent EC2 instance, so you pay standard EC2 rates, plus the usual EKS Auto Mode or ECS Managed Instances management fee if you use them",
      "free_tier": null,
      "compliance": null,
      "regions": "All commercial AWS Regions, on most Nitro-based instance types with Intel, AMD or Graviton processors; flex instances are excluded, as are the smallest and bare-metal sizes of most families",
      "best_for": "AWS teams that want a hardware-enforced VM boundary around each agent session on EKS, ECS or their own EC2 fleet, at no charge beyond the instances",
      "overview": [
        "AWS Nitro Sandboxes, announced on 9 November 2026, is an Amazon EC2 feature rather than a sandbox service. A Nitro Sandbox is a lightweight VM that runs beside an EC2 instance (the parent) on vCPUs and memory the parent contributes. It builds on Nitro Enclaves, with changes for general workloads: vCPUs can time-share with the parent, and the parent attaches virtio network and block devices. Because the sandbox runs directly on the Nitro Hypervisor instead of a nested one, AWS says compute runs at EC2 instance-level performance. Firecracker 1.18 or later is currently the only supported virtual machine monitor.",
        "There are three ways to use it. On Amazon EKS (Kubernetes 1.36 or later), setting `runtimeClassName` to `isolated-sandbox` or `confidential-sandbox` runs that pod in its own sandbox; EKS Auto Mode ships the runtime classes, and self-managed nodes need a sandbox-enabled AMI. On Amazon ECS with Managed Instances, a `trustedExecutionConfiguration` block in the task definition does the same per task. On plain EC2 you launch sandboxes yourself with Firecracker on Amazon Linux 2023, and you own guest images, networking and resource accounting.",
        "Two trust models are offered. In isolated mode (direct mode on EC2) the node is trusted and the goal is to contain the workload. In confidential mode the parent is untrusted: its access to sandbox memory is removed, a Nitro Security Module signs an attestation document, and AWS KMS key policies can release secrets only to a sandbox whose measurements match. AWS names AI agents that run code from their users as a target workload, and points to Lambda MicroVMs for teams that do not want to manage instances. Sandboxes cannot be snapshotted, paused or given GPUs."
      ],
      "notable": [
        "Up to 5 sandboxes per parent vCPU with CPU overcommit, at most 896 per instance, and at least 256 MiB of memory each",
        "On EKS a one-line `runtimeClassName` change sandboxes a pod, and the annotation `sandbox.amazonaws.com/network-isolated` boots it with no network device",
        "Confidential mode adds attestation through a Nitro Security Module, with a Cedar policy on EKS controlling images, exec and logs"
      ],
      "pros": [
        "You already run agents on EKS, ECS or EC2 and want VM isolation per pod, task or session without a new vendor",
        "You want a boundary enforced by the Nitro hypervisor rather than a shared kernel or nested virtualisation",
        "You need to keep workload memory from your own node operators and gate secrets on attestation",
        "You want no charge beyond the EC2 instances you already pay for"
      ],
      "cons": [
        "An EC2 feature rather than a sandbox API: no SDK for creating sessions, running commands or moving files is documented",
        "No snapshots, pause and resume, live migration or GPU passthrough",
        "Sandboxes stop whenever the parent instance stops or reboots, and you size and operate those instances",
        "On plain EC2 you must block the parent's instance metadata service yourself, or a sandbox can read its IAM role credentials"
      ],
      "verdict": "An Amazon EC2 feature that gives EKS pods, ECS tasks and Firecracker workloads Nitro-enforced VM isolation at no additional charge, for teams that run their own instances on AWS.",
      "peers": [
        "aws-lambda-microvms",
        "aws-agentcore",
        "gke-agent-sandbox",
        "kata-containers",
        "firecracker",
        "edera"
      ],
      "sources": [
        "https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/amazon-ec2-nitro-sandboxes.html",
        "https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/nitro-sandboxes-run-firecracker.html",
        "https://docs.aws.amazon.com/eks/latest/userguide/sandboxing.html",
        "https://aws.amazon.com/blogs/compute/nitro-sandboxes-hardware-isolation-for-every-eks-pod-ecs-task-and-ec2-workload/",
        "https://aws.amazon.com/blogs/containers/aws-nitro-sandboxes-hardware-enforced-workload-isolation-on-amazon-eks/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "claude-managed-agents",
      "name": "Claude Managed Agents",
      "company": "Anthropic",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.anthropic.com/legal/privacy"
      },
      "kind": "hyperscaler",
      "tier": 2,
      "alternatives": true,
      "website": "https://platform.claude.com/docs/en/managed-agents/overview",
      "docs": "https://platform.claude.com/docs/en/managed-agents/overview",
      "pricing_url": "https://platform.claude.com/docs/en/about-claude/pricing#claude-managed-agents-pricing",
      "github": null,
      "tagline": "Anthropic's hosted agent harness: each session runs Claude in its own Anthropic-managed Linux container, or sends tool calls to a sandbox you run through a worker.",
      "isolation": "Cloud environments: an isolated Linux container per session on Anthropic-managed infrastructure (the container runtime is not named). Self-hosted sandboxes: bash and file tool calls run wherever your worker runs them, with isolation set by you",
      "isolation_short": "Hosted Linux container",
      "isolation_class": "container",
      "open_source": false,
      "license": null,
      "self_host": true,
      "byoc": true,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "Java",
        "C#",
        "PHP",
        "Ruby",
        "REST API",
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Conversation history is kept until you delete the session. When a session goes idle the service checkpoints its sandbox, including the filesystem, installed packages and files the agent created; sandbox state is kept for 30 days from sandbox creation, after which a resumed session starts from a fresh sandbox. Memory stores mount at /mnt/memory/ and deliverables go to /mnt/session/outputs",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented; cloud sandboxes run Ubuntu 24.04 on x86_64 with up to 8 GB of memory and up to 10 GB of disk",
      "networking": "Set per environment: `limited` allows only hosts in allowed_hosts, plus optional package-manager hosts and the agent's MCP servers; `unrestricted` allows any host except a general safety blocklist. API create requests that omit the field get `unrestricted`; the Console form starts at Limited. With self-hosted sandboxes, network policy is yours",
      "pricing": "Tokens at standard model rates (prompt caching multipliers apply; web search $10 per 1,000 searches) plus $0.08 per session-hour, metered to the millisecond only while the session is `running`; idle, rescheduling and terminated time is not billed. Session runtime replaces code execution container-hour billing. Self-hosted Managed Agents: pricing through the Anthropic account team",
      "free_tier": null,
      "compliance": "Not eligible for zero data retention (except self-hosted Managed Agents) or HIPAA BAA coverage; sessions are stored server-side until deleted",
      "regions": "Beta on the Claude API, Claude Platform on AWS and Claude Models on Google Cloud (Anthropic managed); no sandbox region selection documented. An agent's model.inference_geo can pin inference to the US at 1.1x token rates",
      "best_for": "Long-running Claude agents that need a managed sandbox with network access and persistent session state, without building an agent loop",
      "overview": [
        "Claude Managed Agents is Anthropic's managed agent harness, in beta behind the `managed-agents-2026-04-01` header. You define an agent (model, system prompt, tools, MCP servers, skills), an environment, and then start sessions and exchange events over server-sent events. In a cloud environment each session gets its own fresh Linux container on Anthropic-managed infrastructure: Ubuntu 24.04 on x86_64 with up to 8 GB of memory and 10 GB of disk, preinstalled Python, Node.js, Go, Rust, Java, Ruby, PHP and C/C++ toolchains, PostgreSQL and Redis, and Playwright with Chromium. Networking is `limited` to an allowlist or `unrestricted`, and environments can install apt, cargo, gem, go, npm and pip packages before the agent starts. Built-in tools are bash, file operations, web search and fetch, and MCP servers.",
        "This differs from the Claude code execution tool, which is a single server-side tool added to one Messages API request: its container has 1 CPU and 5 GiB of RAM, no internet access, and is billed per container-hour. Managed Agents instead runs the whole agent loop server-side for minutes or hours, with network access, idle checkpoints of the sandbox and billing of $0.08 per session-hour while running, in place of container-hours. Tool calls can also run in a self-hosted sandbox, where a worker on your infrastructure (prebuilt in the `ant` CLI and the Python, TypeScript and Go SDKs) claims work and runs bash and file calls locally; Anthropic's docs list guides from sandbox providers including E2B, Daytona, Modal, Cloudflare and Vercel. Running the whole service on your own infrastructure (self-hosted Managed Agents) is available by contacting the Anthropic account team. Maximum session length, boot time and GPU support are not documented."
      ],
      "notable": [
        "Session runtime is $0.08 per hour, metered to the millisecond only while a session is `running`, and replaces code execution container-hour billing.",
        "Idle sessions have their sandbox checkpointed; sandbox state is kept for 30 days from creation regardless of activity.",
        "Anthropic's docs list self-hosted sandbox guides from E2B, Daytona, Modal, Cloudflare, Vercel, Fly.io, Blaxel and others."
      ],
      "pros": [
        "You want Anthropic to run the agent loop, context management and sandbox for long-running Claude tasks.",
        "You need a sandbox with network access and preinstalled toolchains, controlled by a host allowlist.",
        "You want to keep files and processes on your own infrastructure while Anthropic runs the agent, via self-hosted sandboxes.",
        "You want runtime billed only while the agent is actively working."
      ],
      "cons": [
        "Beta, and tied to Claude models; it is not a general-purpose sandbox API for your own code.",
        "Not eligible for zero data retention (except self-hosted Managed Agents) or HIPAA coverage.",
        "Maximum session length, boot time, user-triggered snapshots and GPU support are not documented.",
        "API-created environments default to `unrestricted` networking unless you set `limited`."
      ],
      "verdict": "Anthropic's managed agent runtime with a per-session container, for teams that want the agent loop and sandbox run for them rather than a raw sandbox API.",
      "peers": [
        "anthropic-code-execution",
        "aws-agentcore",
        "openai-code-interpreter",
        "e2b",
        "daytona",
        "gke-agent-sandbox"
      ],
      "sources": [
        "https://platform.claude.com/docs/en/managed-agents/overview",
        "https://platform.claude.com/docs/en/managed-agents/choosing-an-environment",
        "https://platform.claude.com/docs/en/managed-agents/environments",
        "https://platform.claude.com/docs/en/managed-agents/cloud-sandboxes-reference",
        "https://platform.claude.com/docs/en/managed-agents/self-hosted-sandboxes",
        "https://platform.claude.com/docs/en/managed-agents/self-hosted",
        "https://platform.claude.com/docs/en/managed-agents/events-and-streaming",
        "https://platform.claude.com/docs/en/managed-agents/session-operations",
        "https://platform.claude.com/docs/en/about-claude/pricing",
        "https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool",
        "https://www.anthropic.com/legal/privacy"
      ],
      "verified_on": "2026-11-18"
    },
    {
      "id": "browserbase",
      "name": "Browserbase",
      "company": "Browserbase",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.browserbase.com/terms-of-service"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.browserbase.com",
      "docs": "https://docs.browserbase.com",
      "pricing_url": "https://www.browserbase.com/pricing",
      "github": "https://github.com/browserbase",
      "tagline": "Hosted headless Chromium sessions for web agents, driven over CDP with Playwright, Puppeteer, Selenium or the Stagehand SDK, with Contexts for persisted logins and four regions.",
      "isolation": "Each session is an isolated Chromium instance with its own user data directory, fingerprint and network identity; the enterprise security page states each browser runs in a dedicated VM that is killed and recreated after each session",
      "isolation_short": "Dedicated VM per browser",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "REST API",
        "MCP server"
      ],
      "cold_start": "Vendor blog claims a session is ready to be driven in under 2 seconds",
      "cold_start_short": "<2 s (claim)",
      "max_session": "Maximum 6 hours per session; Free plan sessions are limited to 15 minutes; keepAlive (paid plans) keeps a session running after disconnect until its timeout or explicit release",
      "max_session_short": "6 h (15 min Free)",
      "persistence": "Sessions start fresh by default; Contexts persist cookies, auth, localStorage, IndexedDB and other profile data across sessions (not the HTTP cache) and are encrypted at rest",
      "snapshots": null,
      "pause_resume": null,
      "gpu": false,
      "gpu_detail": "No — not offered",
      "networking": "Managed proxies with rotation and geo-targeting or bring your own proxy; allowed-domains setting; captcha solving on by default; automatic fingerprinting",
      "pricing": "Free $0: 1 browser hour, 3 concurrent, 15-minute sessions. Developer $20/month: 100 hours, 25 concurrent, $0.12 per extra hour, 1 GB proxy then $12/GB. Startup $99/month: 500 hours, 100 concurrent, $0.10 per extra hour, 5 GB proxy then $10/GB. Scale: custom, 250+ concurrent, usage-based",
      "free_tier": "Free plan: 1 browser hour, 3 concurrent browsers, 15-minute sessions, 7-day data retention",
      "compliance": "SOC 2 Type II; HIPAA BAA and DPA available on request; SSO/SAML",
      "regions": "us-west-2 (default), us-east-1, eu-central-1, ap-southeast-1",
      "best_for": "Web agents that need managed Chromium at scale with persisted logins and stealth",
      "overview": [
        "Browserbase is a hosted browser platform from the company of the same name. You create a session through the REST API or the Node.js and Python SDKs and drive it over the Chrome DevTools Protocol with Playwright, Puppeteer or Selenium, or with Stagehand, Browserbase's own agent SDK. Each session is an isolated Chromium instance with its own user data directory, fingerprint and network identity; the enterprise security page states that each browser runs in a dedicated VM. Sessions can be placed in one of four regions and come with live view, logs and session replay.",
        "Sessions are fresh by default and end on disconnect or timeout (6 hours maximum); **Contexts** persist cookies, auth and site storage between sessions. Pricing is a monthly plan plus browser hours and proxy gigabytes. The platform itself is not open source; the Stagehand SDK and MCP server are open. The company reports 35 million-plus sessions a month across 10,000-plus customers and is SOC 2 Type II audited with HIPAA BAAs on request."
      ],
      "notable": [
        "$40M Series B led by Notable Capital, announced June 2025 (company blog), alongside the Director tool and Stagehand for Python",
        "Homepage states 35M+ browser sessions a month and 10,000+ customers",
        "Four regions: us-west-2, us-east-1, eu-central-1 and ap-southeast-1"
      ],
      "pros": [
        "You want a managed Chromium that works with your existing Playwright, Puppeteer or Selenium code",
        "You need persisted logins across runs via Contexts",
        "You want EU or Asia-Pacific placement and audited compliance (SOC 2 Type II, HIPAA BAA)",
        "You are building with Stagehand or an MCP client and want the hosted server"
      ],
      "cons": [
        "Sessions are capped at 6 hours on self-serve plans, and 15 minutes on the Free plan",
        "Self-hosting and BYOC are not documented; the platform itself is not open source",
        "A general-purpose shell or filesystem sandbox is not documented"
      ],
      "verdict": "A managed-Chromium service for web agents that works with Playwright, Puppeteer, Selenium and Stagehand, with a dedicated VM per browser and Contexts for persisted logins.",
      "peers": [
        "steel",
        "hyperbrowser",
        "anchor-browser",
        "kernel",
        "browser-use-cloud",
        "browserless"
      ],
      "sources": [
        "https://www.browserbase.com",
        "https://www.browserbase.com/pricing",
        "https://www.browserbase.com/enterprise",
        "https://docs.browserbase.com/welcome/introduction",
        "https://docs.browserbase.com/platform/browser/getting-started/create-browser-session",
        "https://docs.browserbase.com/platform/browser/long-sessions/overview",
        "https://docs.browserbase.com/platform/browser/core-features/contexts",
        "https://docs.browserbase.com/optimizations/latency/multi-region",
        "https://docs.browserbase.com/reference/sdk/nodejs",
        "https://docs.browserbase.com/reference/sdk/python",
        "https://docs.browserbase.com/integrations/mcp/introduction",
        "https://browserbase.com/blog/what-is-a-browserbase-browser",
        "https://browserbase.com/blog/series-b-and-beyond",
        "https://docs.browserbase.com/account/enterprise/security",
        "https://docs.browserbase.com/account/billing/plans",
        "https://docs.browserbase.com/platform/browser/long-sessions/timeouts",
        "https://docs.browserbase.com/platform/browser/long-sessions/keep-alive",
        "https://docs.browserbase.com/platform/browser/getting-started/remote-browser-versus-local-browser",
        "https://docs.browserbase.com/platform/functions/overview"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "steel",
      "name": "Steel",
      "company": "Steel",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://steel.dev",
      "docs": "https://docs.steel.dev",
      "pricing_url": "https://steel.dev/pricing",
      "github": "https://github.com/steel-dev/steel-browser",
      "tagline": "Open-source browser API for agents (Apache-2.0, self-hostable with Docker) with a managed Steel Cloud adding stealth, captcha solving, profiles and a hosted MCP server.",
      "isolation": "Each session launches its own isolated Chromium in the cloud; the container or VM boundary is not named. Self-hosted Steel Browser runs as a Docker container",
      "isolation_short": "Isolated Chromium session",
      "isolation_class": null,
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "Rust",
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Sessions last 5 minutes by default (configurable timeout, optional inactivityTimeout); maximum 15 minutes on Launch, 1 hour on Scale, up to 24 hours on Enterprise",
      "max_session_short": "15 min / 1 h / 24 h",
      "persistence": "Each session keeps its own cookies and storage; Profiles snapshot the browser user data directory (cookies, auth, extensions, settings) for reuse via profileId, up to 300 MB, deleted after 30 days unused",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "No — not offered",
      "networking": "Residential and datacenter proxies with country-to-city targeting or bring your own; dedicated IPs ($5/IP/month on Scale); captcha solving; stealth fingerprinting and an Enterprise-only Stealth Browser Chromium fork",
      "pricing": "Launch $0 + usage: $0.10 per browser hour, 10 concurrent sessions, proxy $10/GB, captcha $3 per 1k. Scale $250/month + usage: $100 credits/month, $0.08 per browser hour, 100 concurrent, proxy $6/GB, captcha $1 per 1k. Enterprise: custom, 1,000+ concurrent. Billed per minute, rounded up",
      "free_tier": "$30 one-time credits on the Launch plan (no monthly fee, no card required); Steel Browser is free to self-host",
      "compliance": "HIPAA-ready BAA listed on the pricing page for Scale and Enterprise; a trust centre is published behind an access request",
      "regions": null,
      "best_for": "Teams that want an open-source browser API they can run themselves or scale on a managed cloud",
      "overview": [
        "Steel is an open-source browser API for AI agents. Steel Browser, the Apache-2.0 server on GitHub, runs Chromium under Puppeteer and CDP and exposes a sessions API that Playwright, Puppeteer, Selenium or any CDP client can connect to; you can run it with a single `docker run` or deploy to Railway. Steel Cloud is the managed version, with SDKs for TypeScript, Python, Go and Rust, a CLI, and a hosted MCP server at mcp.steel.dev. Each session launches its own isolated Chromium; the docs do not describe the boundary beneath it.",
        "Sessions default to 5 minutes and are capped per plan (15 minutes, 1 hour or 24 hours). Profiles persist a browser's user data directory for later sessions, and dedicated IPs give a stable network identity. The cloud adds captcha solving, managed proxies, a Credentials API, a Files API and advanced stealth that the local build lacks; the docs say the defining difference is concurrency (1 locally versus 10 to 1,000+ on the cloud). Pricing is per browser hour plus proxy and captcha usage."
      ],
      "notable": [
        "Steel Browser has 7.9k GitHub stars; the homepage claims 35M+ sessions and 1.5M+ browser hours served",
        "Launch Week v3 (post dated 29 June 2026) added the Stealth Browser Chromium fork, Dedicated IPs, and Rust and Go SDKs",
        "Hosted MCP server at https://mcp.steel.dev/mcp with OAuth or API-key sign-in"
      ],
      "pros": [
        "You want an Apache-2.0 browser server you can self-host with Docker",
        "You need a managed cloud with the same API when concurrency grows",
        "You want SDKs beyond Node and Python (Go and Rust)",
        "You are connecting an MCP client and want a hosted server with profiles and saved logins"
      ],
      "cons": [
        "Launch plan sessions are capped at 15 minutes; 24-hour sessions need Enterprise",
        "The isolation boundary is not documented; cloud browsers run in a single region (us-east)",
        "Captcha solving, credentials, files and the MCP server are cloud-only",
        "No compliance certification is named on the public pages checked; the trust centre requires an access request"
      ],
      "verdict": "An open-source browser API (Apache-2.0) that can be self-hosted for free, with Steel Cloud adding advanced stealth, captcha solving and higher concurrency.",
      "peers": [
        "browserbase",
        "browserless",
        "hyperbrowser",
        "kernel",
        "notte",
        "lightpanda"
      ],
      "sources": [
        "https://steel.dev",
        "https://steel.dev/pricing",
        "https://steel.dev/faq",
        "https://github.com/steel-dev/steel-browser",
        "https://docs.steel.dev/overview/intro-to-steel",
        "https://docs.steel.dev/overview/sessions-api/overview",
        "https://docs.steel.dev/overview/sessions-api/session-lifecycle",
        "https://docs.steel.dev/overview/profiles-api/overview",
        "https://docs.steel.dev/overview/self-hosting/steel-local-vs-steel-cloud",
        "https://docs.steel.dev/overview/mcp",
        "https://steel.dev/blog/launch-week-v3-recap",
        "https://docs.steel.dev/overview/sessions-api/configuration",
        "https://trust.steel.dev"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "hyperbrowser",
      "name": "Hyperbrowser",
      "company": "Hyperbrowser",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/hyperbrowser"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.hyperbrowser.ai",
      "docs": "https://hyperbrowser.ai/docs",
      "pricing_url": "https://www.hyperbrowser.ai/pricing",
      "github": "https://github.com/hyperbrowserai",
      "tagline": "Credit-billed cloud Chrome sessions with profiles, stealth and captcha solving, plus one-shot agent runs (Claude Computer Use, OpenAI CUA, Browser Use) and separate VM-based code sandboxes.",
      "isolation": "Sessions are described as isolated browser instances running in the cloud; the boundary is not named. The separate Sandboxes product states commands run in isolated virtual machines",
      "isolation_short": "Isolated browser instance",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": "Homepage claims under 500 ms from request to a ready browser",
      "cold_start_short": "<500 ms (claim)",
      "max_session": "Team default timeout (configurable in settings); per-session timeoutMinutes from 1 to 720 (12 hours); keepAlive keeps a session running after CDP disconnect until timeout or stop",
      "max_session_short": "Up to 12 h",
      "persistence": "Profiles snapshot the user data directory (cookies, local storage, session storage, cache) and reload it in later sessions; persistChanges controls writes; profiles.fork() copies a profile",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented for browsers; GPU (CUDA) sandboxes exist in the separate Sandboxes product",
      "networking": "Residential proxies at $10/GB with country/state/city selection or custom proxies; captcha solving; basic and ultra stealth modes",
      "pricing": "Credits, 1,000 credits = $1. Browser $0.10/hour, proxy $10/GB, sandboxes $0.045/vCPU-hour + $0.015/GiB-hour, HyperAgent/Browser Use $0.02/step. Free: 5,000 credits, 1 concurrent. Startup $30/month: 30,000 credits, 25 concurrent. Scale $100/month: 100,000 credits, 100 concurrent. Enterprise: custom, 1,000+ concurrent",
      "free_tier": "Free plan: 5,000 credits (about 50 browser hours at base rate), 1 concurrent session, 7-day retention",
      "compliance": "Security programme aligned with SOC 2 Trust Services Criteria with independent assessments; the trust centre lists a SOC 2 Type 1 report on request; production data stored in US cloud databases unless agreed otherwise",
      "regions": "us-central, us-east, us-west, europe-west; asia-south on Enterprise/Select plans",
      "best_for": "Agents that need cheap browser hours with a built-in path to run computer-use models as one-shot tasks",
      "overview": [
        "Hyperbrowser is a cloud platform for AI agents from the Y Combinator-backed company of the same name. Its browser product gives you Chrome sessions you drive with Playwright or Puppeteer over CDP from the Node.js or Python SDK, with live view, recordings, profiles, stealth and captcha solving. The docs describe sessions as isolated browser instances but do not name the boundary. An Agents API runs a task to completion with HyperAgent, Browser Use, Claude Computer Use or OpenAI CUA, either blocking or by polling a job.",
        "Everything bills in credits (1,000 per dollar): browsers at $0.10 per hour, proxies at $10/GB and agent steps at $0.02. Per-session timeouts run from 1 to 720 minutes, and keepAlive holds a session across disconnects. A separate Sandboxes product (CPU, GPU and beta macOS) runs code in isolated virtual machines with volumes and custom Docker images, priced per vCPU-hour and GiB-hour; this record covers the browser product. An `hyperbrowser-mcp` server exposes web fetch, search, crawl, extract and browser-task tools."
      ],
      "notable": [
        "Y Combinator company (YC profile lists Summer 2021 batch, founded 2021, team of 4, San Francisco)",
        "Homepage claims 100M+ browser sessions run and 1,000+ concurrent isolated browsers",
        "Five regions, with asia-south limited to Enterprise and Select plans"
      ],
      "pros": [
        "You want low per-hour browser pricing with a usable free tier",
        "You need to run Claude Computer Use, OpenAI CUA or Browser Use as a hosted task",
        "You want profiles that persist logins and can be forked",
        "You need a European or Asian region"
      ],
      "cons": [
        "Isolation boundary for browsers is not documented",
        "The security page states alignment with SOC 2 criteria; the trust centre lists a SOC 2 Type 1 report on request",
        "Self-hosting and BYOC are not documented"
      ],
      "verdict": "A credit-priced browser cloud with built-in runs for Claude Computer Use, OpenAI CUA and Browser Use; the isolation boundary for browser sessions is not documented.",
      "peers": [
        "browserbase",
        "steel",
        "anchor-browser",
        "browser-use-cloud",
        "kernel",
        "notte"
      ],
      "sources": [
        "https://www.hyperbrowser.ai",
        "https://www.hyperbrowser.ai/pricing",
        "https://www.hyperbrowser.ai/security",
        "https://hyperbrowser.ai/docs",
        "https://www.hyperbrowser.ai/docs/sessions/create",
        "https://hyperbrowser.ai/docs/sessions/lifecycle.md",
        "https://hyperbrowser.ai/docs/sessions/profiles",
        "https://hyperbrowser.ai/docs/sessions/regions.md",
        "https://hyperbrowser.ai/docs/agents/claude-computer-use.md",
        "https://hyperbrowser.ai/docs/sandboxes/introduction.md",
        "https://hyperbrowser.ai/docs/integrations/model-context-protocol.md",
        "https://www.ycombinator.com/companies/hyperbrowser",
        "https://trust.hyperbrowser.ai"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "anchor-browser",
      "name": "Anchor Browser",
      "company": "Anchor",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://anchorbrowser.io",
      "docs": "https://docs.anchorbrowser.io",
      "pricing_url": "https://anchorbrowser.io/#pricing",
      "github": null,
      "tagline": "Cloud browsers for agents with a dedicated VM per browser, managed authentication identities, persistent volumes and a Web-Bot-Auth-verified Chromium fork.",
      "isolation": "Dedicated, isolated virtual machine for each browser instance; browsers are never reused between sessions or customers; hypervisor not named",
      "isolation_short": "Dedicated VM per browser",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": true,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "REST API",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Default max_duration 180 minutes and idle_timeout 5 minutes, both configurable per session; no documented upper bound on max_duration; idle timeout can be disabled",
      "max_session_short": "3 h default, configurable",
      "persistence": "Identities store credentials, MFA methods and the browser profile from the last sign-in so sessions start logged in; Volumes are team-owned folders that outlive sessions for uploads and downloads; profiles via SDK",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Built-in \"Anchor VPN\" rather than third-party proxies; custom proxies and geolocation selection on paid plans; automated captcha bypass; Cloudflare-verified browser agents; Web-Bot-Auth",
      "pricing": "Credit plans: Free $0 (5 credits, up to 5 concurrent, up to 500 sessions); Starter $50/month (50 credits, 25 concurrent); Team $500/month (500 credits, 50 concurrent); Growth $2,000/month (2,000 credits, 200 concurrent); Enterprise custom (500+ concurrent). Overage $1.00 per credit; 0.1 credits per completed task. Calculator lists browser duration $0.09/hour, proxy $8/GB, AI steps $0.01",
      "free_tier": "Free plan: 5 credits, up to 5 concurrent browsers, up to 500 browser sessions",
      "compliance": "SOC 2 Type II, ISO 27001, HIPAA (BAA) and GDPR listed; Zero Data Retention mode; browser session data not stored by design",
      "regions": "AWS us-east-2 (Ohio) by default; EU residency in AWS eu-central-1 (Frankfurt) on Growth; custom regions on Enterprise",
      "best_for": "Authenticated enterprise workflows that need VM-level isolation per browser and managed logins",
      "overview": [
        "Anchor Browser is a cloud browser platform from Anchor (Anchor Forge Inc., a Delaware company), aimed at agents that must sign in to real applications. Every browser runs in its own dedicated virtual machine that is never reused between sessions or customers (the hypervisor is not named). You control sessions over CDP with Playwright or Puppeteer, or through the REST API and the `anchorbrowser` TypeScript and Python SDKs, and an MCP server exists in hosted and open-source forms. Anchor Chromium is a stealth fork, and Cloudflare and 1Password are listed as partners.",
        "The distinctive layer is managed authentication: **Identities** hold credentials, MFA secrets and the last good browser profile, so a session opens already signed in and re-runs the login flow when it expires; **Volumes** are persistent folders shared across sessions. Sessions default to a 3-hour cap and a 5-minute idle timeout, both adjustable. Pricing is credit-based with monthly tiers from $50 to $2,000 plus $1 per overage credit. SOC 2 Type II, ISO 27001, HIPAA and GDPR are listed, with EU residency on Growth and bring-your-own-cloud or on-premise on Enterprise."
      ],
      "notable": [
        "$6M seed round from Blumberg Capital and Gradient Ventures, announced on the company blog on 21 October (no year shown), alongside the b0.dev launch",
        "Hosted in AWS us-east-2 by default with EU residency in eu-central-1 on the Growth plan",
        "Listed on AWS Marketplace; bring-your-own-cloud and on-premise deployment offered on Enterprise"
      ],
      "pros": [
        "You need a dedicated VM per browser that is never reused",
        "You want managed logins with MFA handled for you",
        "You need audited compliance (SOC 2 Type II, ISO 27001, HIPAA, GDPR) and EU residency",
        "You want to deploy into your own cloud or on-premise"
      ],
      "cons": [
        "Pricing is in credits; the pricing calculator lists browser duration at $0.09 per hour",
        "The pricing page lists SOC 2 Type II, ISO 27001 and GDPR on the $2,000/month Growth plan, and HIPAA with a BAA on Enterprise",
        "The platform is not open source; bring-your-own-cloud and on-premise deployment are listed on Enterprise only",
        "A startup-time figure is not documented"
      ],
      "verdict": "Cloud browsers with a dedicated VM per browser instance, paired with managed authentication for enterprise agents.",
      "peers": [
        "browserbase",
        "kernel",
        "steel",
        "hyperbrowser",
        "browser-use-cloud",
        "notte"
      ],
      "sources": [
        "https://anchorbrowser.io",
        "https://anchorbrowser.io/#pricing",
        "https://docs.anchorbrowser.io/introduction",
        "https://docs.anchorbrowser.io/security.md",
        "https://docs.anchorbrowser.io/advanced/session-timeout.md",
        "https://docs.anchorbrowser.io/advanced/volumes.md",
        "https://docs.anchorbrowser.io/essentials/authenticated-applications.md",
        "https://docs.anchorbrowser.io/sdk-reference/overview.md",
        "https://anchorbrowser.io/blog/announcing-our-6m-seed-round-by-blumberg-capital-gradient-ventures",
        "https://anchorbrowser.io/privacy-policy"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "kernel",
      "name": "Kernel",
      "company": "Kernel",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/kernel"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.kernel.sh",
      "docs": "https://www.kernel.sh/docs/introduction",
      "pricing_url": "https://www.kernel.sh/pricing",
      "github": "https://github.com/kernel/kernel-images",
      "tagline": "Browsers for agents where each Chromium runs in its own Unikraft unikernel VM, with standby that stops billing, profiles, and open-source images under Apache-2.0.",
      "isolation": "Each browser runs in a dedicated Unikraft-based unikernel VM, isolated at the hypervisor level with no shared host kernel; hypervisor not named in the docs",
      "isolation_short": "Unikernel VM per browser",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Python",
        "Go",
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": "Docs state a browser is created in about 30 ms at p50 and 105 ms at p99; the unikernels page claims cold starts under 20 ms",
      "cold_start_short": "~30 ms p50 (claim)",
      "max_session": "Timeout defaults to 60 seconds and starts when the browser enters standby (5 seconds after the last client disconnects); configurable up to 72 hours",
      "max_session_short": "Up to 72 h",
      "persistence": "Standby keeps browser state without billing until a client reconnects; Profiles capture the user data directory (cookies, auth, local storage, tabs) and are saved with save_changes; encrypted per organisation",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Headful browsers with GPU acceleration at $0.0008000016 per second on Start-Up and Enterprise plans",
      "networking": "Managed residential proxy and captcha solver in stealth mode; configurable and bring-your-own proxies on higher plans; egress allowlists cover browser traffic through the proxy",
      "pricing": "Usage per second: headless $0.0000166667/s, headful $0.0001333336/s, headful + GPU $0.0008000016/s; standby and idle pool browsers are not billed. Plans: Free $0 + usage ($5 credits, 5 concurrent), Hobbyist $30/month ($10 credits, 10 concurrent), Start-Up $200/month ($50 credits, 150 concurrent, regions, GPU), Enterprise custom",
      "free_tier": "Free plan: $5 monthly credits, 5 concurrent browsers, 1-day replays; images are free to run yourself",
      "compliance": "SOC 2 Type II (report on request); HIPAA; PCI DSS Level 4 (SAQ A); ISO 27001 and GDPR listed as in progress; zero data retention on Enterprise",
      "regions": "us-east (default), eu-west, ap-southeast; region selection on Start-Up and Enterprise plans",
      "best_for": "Agents that need fast, hypervisor-isolated browsers with pay-only-while-active billing",
      "overview": [
        "Kernel is browser infrastructure for web agents from the San Francisco company of the same name (YC S25). Every browser session runs inside a dedicated Unikraft-based unikernel virtual machine, so sessions are isolated at the hypervisor level rather than sharing a host kernel. You drive the Chromium over CDP or WebDriver BiDi with Playwright, through computer-use controls, or from the TypeScript, Python and Go SDKs, the `kernel` CLI and a hosted MCP server. The browser images (`kernel-images`, Apache-2.0) and the VM runtime (`hypeman`, MIT) are on GitHub and can be run locally with Docker.",
        "Five seconds after the last client disconnects a browser enters **standby**, keeping its state without accruing usage; the deletion timeout (60 seconds by default, up to 72 hours) only starts then. Profiles persist the user data directory between browsers. Billing is per second of active browser time, with headful and GPU-accelerated tiers costing more, plus a monthly plan fee from $0 to $200. SOC 2 Type II and HIPAA are listed, and the company raised $22M in seed and Series A funding led by Accel in October 2025."
      ],
      "notable": [
        "$22M seed and Series A led by Accel, announced 15 October 2025; Y Combinator Summer 2025 batch",
        "Browser images open source under Apache-2.0 (kernel-images, 1.1k stars); VM runtime hypeman under MIT",
        "Docs claim ~30 ms p50 browser creation; homepage benchmark claims 794 ms average cold start versus competitors"
      ],
      "pros": [
        "You want each browser in its own VM without paying for idle time",
        "You need long-lived browsers (up to 72 hours) that survive client disconnects",
        "You want to inspect or run the browser images yourself",
        "You need GPU-accelerated headful browsers or regional placement"
      ],
      "cons": [
        "The docs do not say which hypervisor the cloud uses; the unikernel build in kernel-images deploys to Unikraft Cloud, while the Docker image runs locally",
        "Region selection, GPU and BYO proxies require the $200/month Start-Up plan",
        "Published start-time figures measure different things: under 20 ms unikernel cold start, ~30 ms p50 browser creation and 794 ms in a cited third-party benchmark",
        "ISO 27001 and GDPR are listed as in progress on the security page"
      ],
      "verdict": "Hypervisor-isolated browsers in Unikraft unikernel VMs with standby billing and open-source images, suited to agents that need a VM per browser.",
      "peers": [
        "anchor-browser",
        "browserbase",
        "steel",
        "hyperbrowser",
        "browser-use-cloud",
        "unikraft-cloud"
      ],
      "sources": [
        "https://www.kernel.sh/",
        "https://www.kernel.sh/pricing",
        "https://www.kernel.sh/docs/introduction",
        "https://www.kernel.sh/docs/info/unikernels.md",
        "https://www.kernel.sh/docs/info/pricing.md",
        "https://www.kernel.sh/docs/security.md",
        "https://www.kernel.sh/docs/browsers/termination.md",
        "https://www.kernel.sh/docs/browsers/profiles.md",
        "https://www.kernel.sh/docs/browsers/regions.md",
        "https://www.kernel.sh/blog/series-a-announcement",
        "https://github.com/kernel/kernel-images",
        "https://github.com/kernel/hypeman"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "browser-use-cloud",
      "name": "Browser Use Cloud",
      "company": "Browser Use",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/browser-use"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://browser-use.com",
      "docs": "https://docs.browser-use.com",
      "pricing_url": "https://browser-use.com/pricing",
      "github": "https://github.com/browser-use/browser-use",
      "tagline": "Stealth cloud Chromium at $0.06 per browser hour from the makers of the MIT-licensed Browser Use agent library, with profiles, live view, an EU deployment and a hosted agent API.",
      "isolation": "Every run gets its own isolated session (a browser, a proxy exit and a model context); the container or VM boundary is not named",
      "isolation_short": "Isolated browser per run",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Standalone browser sessions default to 60 minutes with a maximum of 240 minutes; V4 agent browsers stop about 2 minutes after the last run (keepAlive 120 to 14,400 seconds) with a hard cap of 4 hours",
      "max_session_short": "1 h default, 4 h max",
      "persistence": "Profiles persist cookies, local storage and login state across browsers via profileId; local logins can be synced into a cloud profile; recording is off by default",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Managed home-IP proxies ($5/GB) with country selection, or your own proxy or direct connection ($0.20/GB); captcha solving included; network policies allow or block domains",
      "pricing": "Pay as you go, $0/month: browsers $0.06 per browser hour billed by the minute (1-minute minimum); home-IP proxy $5/GB; own proxy or direct $0.20/GB; bu-ultrafast agent $0.006 per median task. Concurrency unlocks by lifetime spend: 10 at $0, 50 at $100, 250 at $1,000, 500 at $5,000, 1,000 at $25,000. Enterprise: custom",
      "free_tier": "$1 sign-up credit, one per person, never expires; credits top up from $5",
      "compliance": "SOC 2 Type II (audit period 17 July to 17 October 2025); HIPAA on enterprise plans (not on the EU deployment); signed DPA; zero data retention projects",
      "regions": "US by default; EU deployment in Frankfurt (AWS) with its own API base URL and keys",
      "best_for": "Cheap stealth browser hours for agents built on the Browser Use library or any CDP client",
      "overview": [
        "Browser Use Cloud is the hosted side of Browser Use, whose open-source Python library (MIT, about 120k GitHub stars) is an agent framework rather than a sandbox. The cloud product sells stealth browsers: a hardened Chromium fork with per-session fingerprint randomisation, managed proxies, captcha solving, live preview and optional recording. You create a browser with the `browser-use-sdk` Python or TypeScript package or the REST API and connect over the returned CDP URL with Playwright or Puppeteer. The security page says every run gets its own isolated browser, proxy exit and model context but does not name the boundary.",
        "Browsers bill at $0.06 per hour by the minute, with concurrency tiers unlocked by cumulative spend rather than subscription. Standalone sessions default to 60 minutes and cap at 4 hours; profiles carry cookies and logins between browsers. An EU deployment in Frankfurt keeps browser and agent runs in Europe, though billing and some audit metadata stay in the US. SOC 2 Type II is audited, HIPAA is available on enterprise plans, and a hosted agent API runs tasks per-task or per-token."
      ],
      "notable": [
        "The open-source library is MIT licensed with about 120k GitHub stars; the cloud browsers are a separate paid service",
        "SOC 2 Type II audit by Accorp Partners covering 17 July to 17 October 2025",
        "EU deployment in Frankfurt with a separate API endpoint (api.eu.browser-use.com)"
      ],
      "pros": [
        "You want browser hours billed by the minute at a listed $0.06 per hour",
        "You already use the Browser Use library and want its default cloud browser",
        "You need EU-resident browser runs",
        "You want captcha solving and residential proxies without a monthly fee"
      ],
      "cons": [
        "Isolation boundary is not documented",
        "Sessions cap at 4 hours",
        "Higher concurrency requires cumulative spend ($25,000 for 1,000 browsers)",
        "On-prem deployment is offered to enterprise customers through sales only; the EU deployment disables automatic captcha solving and several agent features"
      ],
      "verdict": "Pay-as-you-go stealth browsers with a SOC 2 Type II audit and an EU deployment, suited to teams already on the Browser Use library.",
      "peers": [
        "browserbase",
        "steel",
        "hyperbrowser",
        "anchor-browser",
        "kernel",
        "notte"
      ],
      "sources": [
        "https://browser-use.com",
        "https://browser-use.com/pricing",
        "https://browser-use.com/security",
        "https://github.com/browser-use/browser-use",
        "https://docs.browser-use.com/cloud/browser/quickstart.md",
        "https://docs.browser-use.com/cloud/browser/stealth.md",
        "https://docs.browser-use.com/cloud/api-v4/browsers/create-browser-session.md",
        "https://docs.browser-use.com/cloud/guides/authentication.md",
        "https://docs.browser-use.com/cloud/guides/concurrency.md",
        "https://docs.browser-use.com/cloud/eu-deployment.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "browserless",
      "name": "Browserless",
      "company": "Browserless",
      "hq": {
        "region": "Washington",
        "country": "US",
        "source": "https://www.browserless.io/privacy-policy"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.browserless.io",
      "docs": "https://docs.browserless.io",
      "pricing_url": "https://www.browserless.io/pricing",
      "github": "https://github.com/browserless/browserless",
      "tagline": "Managed headless Chrome, Firefox and WebKit over WebSocket for Puppeteer, Playwright and Selenium, with BrowserQL stealth, REST endpoints and a Docker image you can self-host.",
      "isolation": "The trust page states browser automation runs in isolated, containerised environments (the per-session boundary is not described); Enterprise offers network isolation and dedicated resources; the self-hosted image runs headless browsers inside Docker",
      "isolation_short": "Not documented",
      "isolation_class": "container",
      "open_source": true,
      "license": "SSPL-1.0 OR Browserless Commercial License",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Python",
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Maximum session time per plan: 2 minutes (Free), 15 minutes (Prototyping), 30 minutes (Starter), 60 minutes (Scale); Enterprise not listed",
      "max_session_short": "2 min to 60 min by plan",
      "persistence": "Persisted sessions and replays retained 1 to 90 days by plan; reconnects count as new connections; Authenticated Profiles save cookies, localStorage and IndexedDB for reuse across sessions",
      "snapshots": null,
      "pause_resume": null,
      "gpu": true,
      "gpu_detail": "GPU-enabled infrastructure listed for Enterprise only",
      "networking": "Residential proxies (6 units/MB) and datacenter proxies (2 units/MB); bring-your-own proxies from Prototyping; automatic captcha solving; stealth routes",
      "pricing": "Units: 1 unit per 30 seconds of browser time per connection; captcha solve 10 units. Free $0: 1k units, 2 concurrent, 2-minute sessions. Prototyping $25/month (annual): 20k units, 10 concurrent, 15-minute sessions, $0.0020 per extra unit. Starter $140/month: 180k units, 40 concurrent, 30-minute sessions, $0.0017 overage. Scale $350/month: 500k units, 100 concurrent, 60-minute sessions, $0.0015 overage. Enterprise: custom, private deployment or self-hosting licence",
      "free_tier": "Free plan: 1,000 units, 2 concurrent browsers, 2-minute sessions; open-source Docker image free for non-commercial use",
      "compliance": "SOC 2 Type II; GDPR; HIPAA-compatible deployments with BAA on Enterprise",
      "regions": "Shared fleet: US West (SFO, default), Europe UK (LON), Europe Amsterdam (AMS); private fleet adds US East (NYC)",
      "best_for": "Existing Puppeteer, Playwright or Selenium automations that want a managed or self-hosted browser pool",
      "overview": [
        "Browserless is a long-running browser-as-a-service from the company of the same name, founded by Joel Griffith. You point Puppeteer, Playwright or Selenium at a WebSocket endpoint in one of three shared-fleet regions, call one-shot REST endpoints for screenshots, PDFs and scraping, or use BrowserQL, a GraphQL layer with managed stealth and captcha solving that the TypeScript and Python BAP SDKs wrap. The core runs in Docker (`ghcr.io/browserless/chromium`) under SSPL-1.0 or a commercial licence, so you can self-host, though BrowserQL, stealth, captcha solving and recording are Enterprise-image features.",
        "The trust page says browser automation runs in isolated, containerised environments without describing the per-session boundary; Enterprise private deployments promise network isolation and dedicated resources. Billing is in units of 30 seconds of browser time per connection, with per-plan concurrency and session caps from 2 minutes on Free to 60 minutes on Scale. An MCP server exposes scraping, screenshots, PDFs and custom Puppeteer code. The trust centre lists SOC 2 Type II and GDPR, with HIPAA-compatible deployments on Enterprise."
      ],
      "notable": [
        "GitHub repository has 13.9k stars; the about page cites 175M+ Docker image pulls and 10,000+ paying customers",
        "Licensed SSPL-1.0 OR Browserless Commercial License; commercial use of the self-hosted image needs the commercial licence",
        "Three shared-fleet regions (SFO, LON, AMS) plus NYC for private fleets"
      ],
      "pros": [
        "You want to keep existing Puppeteer, Playwright or Selenium code unchanged",
        "You need Firefox and WebKit as well as Chrome",
        "You want a Docker image you can run on your own infrastructure",
        "You need European endpoints and SOC 2 Type II"
      ],
      "cons": [
        "The trust page describes containerised browser environments but not the per-session boundary",
        "Sessions are capped at 2 minutes on Free and 60 minutes at most on self-serve plans",
        "SSPL licence restricts commercial self-hosting without a paid licence"
      ],
      "verdict": "A long-running managed browser pool for Puppeteer, Playwright and Selenium automation code, also available as a self-hosted Docker image.",
      "peers": [
        "steel",
        "browserbase",
        "lightpanda",
        "cloudflare-browser-rendering",
        "hyperbrowser",
        "kernel"
      ],
      "sources": [
        "https://www.browserless.io",
        "https://www.browserless.io/pricing",
        "https://www.browserless.io/about",
        "https://www.browserless.io/trust",
        "https://github.com/browserless/browserless",
        "https://docs.browserless.io/",
        "https://docs.browserless.io/overview/connection-urls",
        "https://docs.browserless.io/enterprise/overview.md",
        "https://docs.browserless.io/enterprise/open-source.md",
        "https://docs.browserless.io/bap.md",
        "https://docs.browserless.io/mcp/overview.md",
        "https://docs.browserless.io/baas/features/authenticated-profiles",
        "https://docs.browserless.io/baas/session-management/persisting-state"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "lightpanda",
      "name": "Lightpanda",
      "company": "Lightpanda",
      "hq": {
        "city": "Paris",
        "country": "FR",
        "source": "https://lightpanda.io/legal"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://lightpanda.io",
      "docs": "https://lightpanda.io/docs",
      "pricing_url": "https://lightpanda.io/pricing",
      "github": "https://github.com/lightpanda-io/browser",
      "tagline": "A headless browser written from scratch in Zig (AGPL-3.0) with no graphical rendering, offered as a CDP endpoint on Lightpanda Cloud with a free 10-hour tier.",
      "isolation": "Not documented for Lightpanda Cloud; locally Lightpanda runs as a single headless browser process or Docker container",
      "isolation_short": "Not documented",
      "isolation_class": null,
      "open_source": true,
      "license": "AGPL-3.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "A CDP connection to the cloud closes after 15 minutes on every plan; MCP sessions close after 5 minutes without a request",
      "max_session_short": "15 min per connection",
      "persistence": "None documented; the cloud console lists recent sessions and logs",
      "snapshots": null,
      "pause_resume": null,
      "gpu": false,
      "gpu_detail": "No — the engine has no graphical rendering",
      "networking": "Managed proxy pool with country selection on the cloud, or bring your own proxy; local builds can block private networks and CIDRs",
      "pricing": "Explorer $0: 10 browser hours per month, 5 concurrent sessions. Builder $19/month: 300 hours then $0.08 per hour, 30 concurrent sessions. Enterprise: custom volume pricing, on-premise or private cloud. Usage counted by the second",
      "free_tier": "Explorer plan: 10 browser hours per month, 5 concurrent sessions; the browser is free to self-host",
      "compliance": null,
      "regions": "euwest (west Europe) and uswest (west US) cloud endpoints",
      "best_for": "High-volume page fetching and crawling where memory and speed matter more than pixel-accurate rendering",
      "overview": [
        "Lightpanda is a headless browser built from scratch in Zig by the French company of the same name; it is not a Chromium fork and does not render graphics. It exposes CDP and WebDriver BiDi so Puppeteer, Playwright and chromedp scripts work against it, plus a built-in MCP server and agent in the local build. The README claims it is 16x lighter and 9x faster than headless Chrome on a 933-page benchmark. Version 1.0 (2 October 2026) took it out of beta with 1.74 million passing Web Platform Tests subtests, about 80% of Chrome's count.",
        "Lightpanda Cloud serves the same engine over `wss://euwest.cloud.lightpanda.io/ws` or the US-west equivalent; append `browser=chrome` for a debugging Chrome that is not meant for production. Plans are Explorer (free, 10 hours a month) and Builder ($19 a month for 300 hours), counted by the second, with a 15-minute cap on each CDP connection. The docs do not describe how cloud sessions are isolated, and profile or login persistence is not documented for the cloud. The browser is AGPL-3.0, so self-hosting is free but copyleft."
      ],
      "notable": [
        "Lightpanda 1.0 released 2 October 2026 after two years and 10,000 commits (company blog)",
        "Pre-seed round led by ISAI with Kima, Factorial Capital and Prototype Capital, announced 10 June 2025 (amount undisclosed)",
        "GitHub repository has 41.7k stars under AGPL-3.0"
      ],
      "pros": [
        "You want a lightweight engine for crawling and scraping at volume",
        "You need a free tier with real browser hours",
        "You want to self-host an open-source browser without Chromium",
        "You are running CDP scripts and only need a WebSocket endpoint"
      ],
      "cons": [
        "No graphical rendering, so screenshots return a placeholder image rather than a real render",
        "Passes about 80% of Chrome's Web Platform Tests subtest count (1.0 release post)",
        "Cloud connections are capped at 15 minutes and isolation is not documented",
        "AGPL-3.0 licence for self-hosting"
      ],
      "verdict": "A lightweight headless engine for machine browsing without graphical rendering, self-hostable under AGPL-3.0 or used through Lightpanda Cloud's CDP endpoint.",
      "peers": [
        "browserless",
        "steel",
        "cloudflare-browser-rendering",
        "browserbase",
        "notte"
      ],
      "sources": [
        "https://lightpanda.io",
        "https://lightpanda.io/pricing",
        "https://github.com/lightpanda-io/browser",
        "https://lightpanda.io/docs/run-on-lightpanda-cloud/getting-started",
        "https://lightpanda.io/docs/run-on-lightpanda-cloud/limits-and-billing",
        "https://lightpanda.io/docs/core-concepts/local-vs-cloud",
        "https://lightpanda.io/blog/posts/lightpanda-1-0",
        "https://lightpanda.io/blog/posts/lightpanda-raises-preseed",
        "https://lightpanda.io/docs/core-concepts/architecture-overview"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "cloudflare-browser-rendering",
      "name": "Cloudflare Browser Run",
      "company": "Cloudflare",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.cloudflare.com/about-overview/"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://developers.cloudflare.com/browser-run/",
      "docs": "https://developers.cloudflare.com/browser-run/",
      "pricing_url": "https://developers.cloudflare.com/browser-run/pricing/",
      "github": null,
      "tagline": "Headless Chrome on Cloudflare's network (formerly Browser Rendering), driven from a Workers binding, a REST API or a CDP endpoint, billed by browser hour on the Workers plan.",
      "isolation": "The container or VM boundary is not documented; sessions draw from a global pool of warm Chrome instances, and the FAQ says a fresh browser gives process-level isolation while incognito contexts isolate cookies and storage",
      "isolation_short": "Not documented",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": false,
      "sdks": [
        "JavaScript/TypeScript",
        "REST API",
        "MCP server"
      ],
      "cold_start": "Blog states sessions come from a global pool of warm instances with no cold start waiting for a browser; no figure given",
      "cold_start_short": "Warm pool (claim)",
      "max_session": "Browser closes after 60 seconds of inactivity by default; keep_alive extends the inactivity timeout to at most 20 minutes",
      "max_session_short": "60 s idle, 20 min keep_alive",
      "persistence": "None documented beyond reusing a live browser across requests (for example from a Durable Object); session recordings for debugging",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Runs on Cloudflare's edge near users; no proxy or stealth features documented",
      "pricing": "Workers Free: 10 minutes of browser time per day, 3 concurrent browsers. Workers Paid: 10 browser hours per month included then $0.09 per additional hour; 10 concurrent browsers (averaged monthly) included then $2.00 per additional browser; up to 200 concurrent per account",
      "free_tier": "Workers Free: 10 minutes of browser time per day, 3 concurrent browsers, 1 new browser every 20 seconds",
      "compliance": null,
      "regions": "Cloudflare global network; sessions open near the user",
      "best_for": "Screenshots, scraping and short agent browsing from Workers without running a browser fleet",
      "overview": [
        "Browser Run, renamed from Browser Rendering on 15 April 2026, runs headless Chrome on Cloudflare's global network. Quick Actions are single HTTP requests for screenshots, PDFs, Markdown, structured data and crawls; Browser Sessions give direct control through Cloudflare's forks of Puppeteer and Playwright inside a Worker, through Stagehand, or from any environment over a CDP endpoint added in April 2026. MCP clients can also use it as a remote browser. Cloudflare's docs do not name the container or VM boundary around a browser.",
        "Browsers time out after 60 seconds of inactivity, extendable to 20 minutes with keep_alive, and a Durable Object can hold a live browser across requests; profile or login persistence is not documented. The rename brought Live View, Human in the Loop hand-off, session recordings and WebMCP support. Pricing follows the Workers plan: 10 minutes a day free, or 10 hours a month and 10 concurrent browsers included on Workers Paid, then $0.09 per hour and $2 per extra concurrent browser."
      ],
      "notable": [
        "Generally available on 2 April 2024; renamed Browser Run on 15 April 2026 with Live View, recordings and WebMCP",
        "CDP exposed as an endpoint for external clients on 10 April 2026",
        "Workers Paid concurrency raised from 30 to 120 at the rename; the limits page now lists 200 per account"
      ],
      "pros": [
        "You already run on Cloudflare Workers and want a browser binding",
        "You need one-shot screenshot, PDF or scrape endpoints without infrastructure",
        "You want edge-local sessions near end users",
        "You want a free daily allowance for light use"
      ],
      "cons": [
        "The container or VM boundary around a browser is not documented",
        "Browsers close after 60 seconds of inactivity by default, extendable to 20 minutes with keep_alive",
        "Proxies, stealth, captcha solving and persisted profiles are not documented; the FAQ says requests are always identified as bot traffic",
        "The Free plan allows 3 concurrent browsers and 1 new browser every 20 seconds"
      ],
      "verdict": "Edge-hosted headless Chrome for Workers users and quick page tasks, with browsers that close after 60 seconds of inactivity by default.",
      "peers": [
        "browserless",
        "lightpanda",
        "browserbase",
        "cloudflare-sandbox",
        "steel"
      ],
      "sources": [
        "https://developers.cloudflare.com/browser-run/",
        "https://developers.cloudflare.com/browser-run/limits/",
        "https://developers.cloudflare.com/browser-run/pricing/",
        "https://developers.cloudflare.com/browser-run/get-started/",
        "https://developers.cloudflare.com/browser-run/changelog/",
        "https://blog.cloudflare.com/browser-run-for-ai-agents",
        "https://developers.cloudflare.com/browser-run/faq/",
        "https://developers.cloudflare.com/browser-run/features/reuse-sessions/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "notte",
      "name": "Notte",
      "company": "Notte Labs",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/notte"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://notte.cc",
      "docs": "https://docs.notte.cc",
      "pricing_url": "https://notte.cc/pricing",
      "github": "https://github.com/nottelabs/notte",
      "tagline": "Cloud browser sessions at $0.01 per hour with stealth, proxies, profiles, secret vaults and personas, a hosted MCP server and serverless browser functions; SSPL-licensed core.",
      "isolation": "Sessions are described as isolated browser instances running in the cloud; the container or VM boundary is not named",
      "isolation_short": "Isolated browser instance",
      "isolation_class": null,
      "open_source": true,
      "license": "SSPL-1.0",
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "CLI",
        "REST API",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "max_duration_minutes defaults to 15 (API maximum 1,440) with an idle timeout of 3 minutes (max 30); plan caps are 15 minutes (Free), 120 minutes (Developer), 300 minutes (Startup), custom (Enterprise)",
      "max_session_short": "15 min to 5 h by plan",
      "persistence": "Profiles snapshot cookies, local storage and session storage (not cache, history or extensions) and reload with persist true/false; vaults hold credentials that are never forwarded to LLMs; personas provide email, phone and 2FA identities",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Residential proxies at $10/GB, bring your own at $0.20/GB on Startup; captcha solving and anti-detection; secret proxies keep real credential values away from the browser",
      "pricing": "Browser sessions $0.01 per hour; browser functions $0.05 per hour; residential proxy $10/GB; LLM tokens at cost. Free $0: $10 one-time credits, 5 concurrent, 15-minute sessions. Developer $20/month: $20 credits, 25 concurrent, 120-minute sessions. Startup $100/month: $100 credits, 100 concurrent, 300-minute sessions, BYO keys and proxy. Enterprise: custom, SSO, SLAs, HIPAA, on-premise",
      "free_tier": "Free plan: $10 credits granted once with no expiry, 5 concurrent browsers, 15-minute sessions, 5 vaults and 5 personas",
      "compliance": "SOC 2 Type II; HIPAA and on-premise deployment on Enterprise",
      "regions": null,
      "best_for": "Agents that need cheap browser sessions with credential vaults and synthetic identities for sign-ups and logins",
      "overview": [
        "Notte is a browser platform for AI agents from Notte Labs, a Y Combinator W25 company. Cloud sessions are isolated browser instances with anti-detection, proxies and captcha solving built in; you attach Playwright over the session's CDP URL, use the `notte-sdk` Python or Node packages, the Homebrew CLI, or the hosted MCP server at api.notte.cc/mcp, which exposes browser observe, action, scrape, file and profile tools. The docs do not say what sits beneath a session. Browser Functions deploy automations as serverless endpoints with cron scheduling.",
        "Sessions default to 15 minutes with plan caps up to 5 hours on Startup; profiles persist cookies and storage, vaults keep credentials out of model context and personas supply email, phone and 2FA identities. Browser time is $0.01 per hour on top of monthly credit plans from $0 to $100. The core repository is open under SSPL-1.0 (local Playwright sessions and scraping), while stealth sessions, functions, vaults and personas are hosted only. SOC 2 Type II is stated, with HIPAA and on-premise on Enterprise."
      ],
      "notable": [
        "Y Combinator W25; SOC 2 Type II stated on the homepage",
        "Core repository licensed SSPL-1.0 with 2.1k GitHub stars",
        "Maintains BrowserArena, an open-source benchmark comparing cloud browser providers"
      ],
      "pros": [
        "You want the lowest listed browser-hour price with a no-expiry free credit",
        "You need vaults and personas for sign-ups, logins and 2FA",
        "You want a hosted MCP server with profiles and auth tools",
        "You want to schedule automations as serverless browser functions"
      ],
      "cons": [
        "Isolation boundary and regions are not documented",
        "Free and Developer sessions cap at 15 and 120 minutes",
        "SSPL licence on the open-source core; the cloud features are not self-hostable below Enterprise",
        "The SDK agent loop (notte.Agent, client.Agent) was removed in version 1.10; the README points to sessions, the CLI or the MCP server instead"
      ],
      "verdict": "A low-cost browser cloud with credential vaults and persona identities for sign-ups and logins; the isolation boundary is not documented.",
      "peers": [
        "steel",
        "hyperbrowser",
        "browser-use-cloud",
        "anchor-browser",
        "browserbase",
        "kernel"
      ],
      "sources": [
        "https://notte.cc",
        "https://notte.cc/pricing",
        "https://github.com/nottelabs/notte",
        "https://docs.notte.cc/concepts/sessions.md",
        "https://docs.notte.cc/features/sessions/lifecycle.md",
        "https://docs.notte.cc/features/sessions/browser-profiles.md",
        "https://docs.notte.cc/mcp-server.md",
        "https://docs.notte.cc/features/sessions/configuration.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "scrapybara",
      "name": "Scrapybara",
      "company": "Scrapybara",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://scrapybara.com"
      },
      "status": "winding-down",
      "status_label": "docs redirect to Capy",
      "status_note": "Scrapybara’s own site (checked 15 November 2026) says \"We started by building computers for agents. Now we build Capy\" and lists \"Computers for agents\" under 2024 to 2025. Its documentation site, docs.scrapybara.com, redirects to the Capy docs, and it has no public pricing page. We found no shutdown announcement, and its SDK repositories remain on GitHub; we list it as inactive because its documentation is no longer online, not because the company has announced a closure.",
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://scrapybara.com",
      "docs": null,
      "pricing_url": null,
      "github": "https://github.com/scrapybara",
      "tagline": "Ubuntu, browser and Windows desktop instances for computer-use agents with Python and TypeScript SDKs, which the company lists on its own timeline under 2024 to 2025; it now builds Capy.",
      "isolation": "Not documented; the vendor calls each desktop an \"instance\" and does not name a hypervisor or container runtime",
      "isolation_short": "Not documented",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "REST API",
        "MCP server",
        "CLI"
      ],
      "cold_start": "A March 2025 vendor blog post says \"Instances come online in less than 1 second\"; the Scrapybara docs, which covered Windows startup, are no longer online",
      "cold_start_short": "<1 s (claimed)",
      "max_session": "Instances stop automatically after 1 hour by default; a longer `timeout_hours` can be set at launch (no documented maximum)",
      "max_session_short": "1 h default, configurable",
      "persistence": "When the Scrapybara docs were last checked (9 November 2026) they described pause and resume endpoints, said stopping an instance deletes its session data, and described saving browser auth states for reuse; a snapshot feature was not documented. Those docs are no longer online",
      "snapshots": null,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not documented",
      "networking": "Internet access by default; a `blocked_domains` list can be passed when starting an instance",
      "pricing": "No public pricing page (scrapybara.com/pricing returns 404). When last checked (9 November 2026) the docs said instances were \"billed per usage\", with Ubuntu and Browser instances at \"1x compute cost\" and Windows at \"2x compute cost\"; those docs are no longer online",
      "free_tier": null,
      "compliance": null,
      "regions": null,
      "best_for": "Computer-use agents that need a ready-made Ubuntu or Windows desktop with a browser, bash, filesystem and Jupyter tools in one API",
      "overview": [
        "Scrapybara hosts remote desktop instances for computer-use agents. Through the Python or TypeScript SDK (or the REST API) you start an **Ubuntu 22.04** desktop, a **Browser** instance, or a **Windows 11** desktop (early access), then drive it with a `computer` tool for mouse and keyboard actions, a `bash` tool, a `file` tool, built-in Jupyter notebooks and a Playwright/CDP browser protocol. An interactive stream URL lets a human watch or take over. The docs provide ready-made tool classes for Anthropic computer use and OpenAI CUA, and an \"Act SDK\" that runs an agent loop against an instance. The isolation technology is not documented.",
        "Instances stop after a default one-hour timeout unless you set `timeout_hours`; they can be paused and resumed, and browser auth states can be saved and loaded into later instances. There is no public pricing page; when the docs were last checked (9 November 2026) they said instances were billed per usage, with Windows at twice the compute cost of Ubuntu. The company homepage, titled \"Scrapybara, the company behind Capy\", now reads \"We started by building computers for agents. Now we build Capy\", a platform for engineering teams to run agent swarms. As of 15 November 2026 docs.scrapybara.com redirects to the Capy docs; the Scrapybara SDK and MCP repositories remain on GitHub, unarchived and without a deprecation notice."
      ],
      "notable": [
        "Company homepage (checked 15 November 2026) lists \"Computers for agents\" under 2024 to 2025 and Capy under \"Now\"; docs.scrapybara.com redirects to the Capy docs",
        "Python and TypeScript SDKs, an MCP server (MIT) and a CLI are public on GitHub; the source of the hosted service is not published"
      ],
      "pros": [
        "You want Ubuntu, browser and Windows desktops behind one SDK with computer, bash, file and notebook tools built in",
        "You need pre-built tool definitions for Anthropic computer use or OpenAI CUA",
        "You want to save browser auth states and reuse logged-in sessions across instances"
      ],
      "cons": [
        "Isolation technology, pricing, regions and compliance are not documented",
        "The company's site says it now builds Capy, and docs.scrapybara.com redirects to the Capy docs",
        "No snapshot or image feature is documented on the vendor's remaining public pages"
      ],
      "verdict": "A computer-use desktop API (Ubuntu, browser and Windows instances) from Scrapybara, Inc., which says it now builds Capy and whose Scrapybara docs site redirects to the Capy docs; suited only to teams that have confirmed its current availability with the vendor.",
      "peers": [
        "cua",
        "orgo",
        "kasm-workspaces",
        "browserbase",
        "kernel",
        "hyperbrowser"
      ],
      "sources": [
        "https://scrapybara.com",
        "https://github.com/scrapybara",
        "https://capy.ai",
        "https://scrapybara.com/blog",
        "https://scrapybara.com/blog/speed-scale-stability",
        "https://github.com/scrapybara/scrapybara-python",
        "https://github.com/scrapybara/scrapybara-mcp"
      ],
      "verified_on": "2026-11-15"
    },
    {
      "id": "cua",
      "name": "Cua",
      "company": "Cua AI, Inc.",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/cua"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://cua.ai",
      "docs": "https://cua.ai/docs",
      "pricing_url": "https://cua.ai/pricing",
      "github": "https://github.com/trycua/cua",
      "tagline": "Open-source computer-use infrastructure: Linux sandboxes as gVisor containers and macOS or Windows sandboxes as VMs, run locally, in your own cloud account or as shared \"Spaces\".",
      "isolation": "Linux images run as gVisor containers by default (runc optional); macOS and Windows images run as VMs via Lume (Apple Virtualization.Framework on Apple silicon) or QEMU",
      "isolation_short": "gVisor container or VM",
      "isolation_class": "mixed",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": true,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Rust",
        "CLI",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Persistent sandboxes live until you destroy them; ephemeral sandboxes are deleted when the `async with` block exits. Spaces plans meter \"remote hours\" per month (20 Free, 200 Pro)",
      "max_session_short": "Until destroyed",
      "persistence": "suspend, resume and restart keep state: containers are paused or checkpointed, QEMU VMs are snapshotted, Lume VMs are stopped with their disk kept",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not documented in the sandbox reference",
      "networking": "`network=\"default\"` gives outbound access like a Docker container; `network=\"none\"` removes egress but requires a local QEMU VM (unsupported for containers and Lume). Ports exposed via service(), public_url() (TTL 60 s to 24 h) and tunnels",
      "pricing": "SDK, CLI, Driver and Lume are free (MIT). Cua Spaces: Free $0; Pro $18/month or $15/month billed yearly; Teams $30/seat/month or $25/seat/month billed yearly. Cua does not host compute; VMs in your AWS, Google Cloud or Modal account are billed by that provider",
      "free_tier": "Free to self-host; Spaces Free plan includes 2 computers in use at once and 20 remote hours per month",
      "compliance": null,
      "regions": null,
      "best_for": "Teams that want open-source computer-use sandboxes on their own Macs, servers or cloud accounts, including macOS VMs on Apple silicon",
      "overview": [
        "Cua (trycua/cua on GitHub, about 32k stars) is an open-source toolkit for computer-use agents. Its `cua` SDK (Python, TypeScript, Rust) and CLI create sandboxes from published images: Ubuntu 24.04 runs as a **gVisor** container by default, while macOS 15/26, Windows Server 2022, Omarchy and Android images run as VMs through **Lume** (Apple's Virtualization.Framework on Apple silicon) or QEMU, with HVF, KVM or Hyper-V acceleration depending on the host. Sandboxes default to the local machine; the docs also describe running them in your own AWS, Google Cloud or Modal account. **Cua Driver** lets agents operate native apps and browsers on macOS, Windows and Linux via CLI, MCP or SDK, and **Cua Bench** creates and scores computer-use tasks.",
        "Sandboxes persist until destroyed (or are ephemeral within a context block); suspend/resume checkpoints containers and snapshots QEMU VMs. **Cua Spaces** is a desktop app that turns computers you own into shared desktops for agents, with a relay for remote use; it is source-available under FSL-1.1-MIT (each release becomes MIT after two years) and is the paid tier ($18/month Pro, $30/seat Teams). Cua's hosted **Cua Cloud and Fleets ended on 18 October 2026**, and the older `cua-computer`, `cua-agent` and `cuabot` packages are no longer maintained; the migration guide points to local, own-cloud or Spaces targets instead."
      ],
      "notable": [
        "Cua Cloud and Fleets ended on 18 October 2026; `on=\"cloud\"` now raises a \"Cua Cloud has ended\" error (migration guide)",
        "Repository has about 32.3k stars and 2.4k forks; core is MIT, Cua Spaces and Driver Bench are FSL-1.1-MIT, cua-som is AGPL-3.0-or-later",
        "Spaces pricing: Free (2 computers, 20 remote hours/month), Pro $18/month, Teams $30/seat/month"
      ],
      "pros": [
        "You want macOS, Linux and Windows computer-use sandboxes from one open-source SDK",
        "You need macOS VMs on Apple silicon for agents that must drive native Mac apps",
        "You run sandboxes on your own hardware or cloud account and want no vendor-hosted compute",
        "You are building computer-use benchmarks or training data with Cua Bench"
      ],
      "cons": [
        "No hosted compute: you provision and pay for machines yourself (Cua Cloud ended October 2026)",
        "Licensing differs by component: MIT for the core, FSL-1.1-MIT for Cua Spaces and Driver Bench, AGPL-3.0-or-later for the optional cua-som package",
        "Egress-free networking is only supported for local QEMU VMs, not gVisor containers or Lume",
        "No compliance attestations are documented"
      ],
      "verdict": "An open-source toolkit for computer-use sandboxes across macOS, Linux and Windows that you run on your own machines or cloud account, suited to teams that want no vendor-hosted compute; Cua's hosted Cua Cloud ended on 18 October 2026.",
      "peers": [
        "scrapybara",
        "orgo",
        "kasm-workspaces",
        "microsandbox",
        "docker-sandboxes",
        "gvisor"
      ],
      "sources": [
        "https://cua.ai/",
        "https://cua.ai/pricing",
        "https://cua.ai/docs",
        "https://cua.ai/docs/cua-sdk/reference/python/sandbox",
        "https://cua.ai/docs/cua-sdk/guides/cua-cloud-ended",
        "https://github.com/trycua/cua"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "orgo",
      "name": "Orgo",
      "company": "Orgo",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.orgo.ai/terms"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.orgo.ai",
      "docs": "https://docs.orgo.ai",
      "pricing_url": "https://www.orgo.ai/pricing.md",
      "github": null,
      "tagline": "Always-on cloud Linux and Windows desktops for AI agents, each an isolated VM billed by the second, with a computer-use API, MCP server, SSH, clone and fork.",
      "isolation": "Each Linux and Windows computer is \"its own isolated virtual machine with its own disk\"; the docs do not name the hypervisor (the API returns a per-computer `hypervisor` field)",
      "isolation_short": "Isolated VM (hypervisor not named)",
      "isolation_class": "vm",
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "REST API",
        "CLI",
        "MCP server"
      ],
      "cold_start": "Docs say \"a new computer is online in seconds\"; no figure given",
      "cold_start_short": "Seconds (claimed)",
      "max_session": "Paid computers run until you stop or delete them (never auto-suspended); optional end times of 1 to 720 hours, extendable up to 2,160 hours ahead. Free-plan computers stop when credit runs out",
      "max_session_short": "Unlimited (paid)",
      "persistence": "Files and installed software persist across stops; a stopped computer cold-boots from its saved disk (memory not kept). Sleep keeps memory. Clone and fork endpoints and reusable templates; free-plan files kept 7 days after credit runs out",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Linux-only `--gpu` option with 2 GB (`2q`) or 4 GB (`4q`) slices; falls back to CPU if no GPU host is free",
      "networking": "Outbound internet; 1 Gbps network on paid plans; own HTTP CONNECT proxy via the egress API; residential IPs (US-only, Linux-only) with a monthly data allowance on paid plans; public URLs for ports on paid plans; direct SSH for Linux",
      "pricing": "Per second, no hourly minimum: Small (0.5 vCPU, 4 GB) $0.023/h running, $0.0014/h stopped; Medium (2 vCPU, 8 GB, 40 GB) $0.055/h running, $0.0027/h stopped; Large (4 vCPU, 16 GB) $0.11/h; XL (6 vCPU, 32 GB) $0.202/h; 2XL (8 vCPU, 64 GB) $0.368/h. Plans: Pay as you go $0; Starter $29/month (1 Medium always on); Growth $99/month (4 Medium); Scale $399/month (16 Medium, Windows included); Enterprise custom. Windows is $15/month per computer on Starter/Growth",
      "free_tier": "$5 starting credit for new accounts (about 90 running hours of one Medium computer); $2 for agent-created accounts; one computer at a time on Free",
      "compliance": null,
      "regions": null,
      "best_for": "Agents that keep a persistent cloud desktop running around the clock with a browser, terminal and SSH",
      "overview": [
        "Orgo provides cloud computers for AI agents: full Linux and Windows desktops that \"stay on 24/7\", each described as its own isolated virtual machine with its own disk, a desktop, a browser and a terminal, reachable over HTTP and VNC. You control them through a REST API, Python and TypeScript SDKs, a CLI (`npx orgo init`) or a hosted MCP server that plugs into Claude Code, Codex, Cursor and similar clients. The computer-use API exposes click, type, key, scroll, drag, screenshot, bash and Python exec actions, and the docs include integration guides for Claude, OpenAI and Gemini computer-use models. Macs and iPhones are offered separately while in stock.",
        "Computers are billed by the second with no hourly minimum; a Medium (2 vCPU, 8 GB) costs $0.055 per hour running and $0.0027 per hour stopped. Files and installed software persist across stops, and computers can be cloned, forked and relaunched from templates. Subscription plans from $29 per month keep one or more computers always on, and new accounts get $5 of credit. The docs do not name the hypervisor or list hosting regions; a trust centre at trust.orgo.ai lists security controls, and no named compliance attestation is published."
      ],
      "notable": [
        "Per-second billing: Medium computer $0.055/h running, $0.0027/h stopped; $5 starting credit (pricing page)",
        "Linux GPU slices of 2 GB or 4 GB are available; Windows costs $15/month per computer on lower plans",
        "Hosted MCP server at orgo.ai/mcp and agent self-signup accounts ($2 credit, expire after 72 hours unless claimed)"
      ],
      "pros": [
        "You want an agent to keep a persistent desktop that survives stops and restarts",
        "You need Windows as well as Linux desktops, and SSH access to the Linux ones",
        "You want to connect an MCP client such as Claude Code directly to a cloud computer",
        "You are watching cost: stopped computers cost fractions of a cent per hour"
      ],
      "cons": [
        "Hypervisor and hosting regions are not documented; the trust centre lists security controls but no named compliance attestation",
        "Residential egress is US-only and Linux-only, with no data included on the Free plan",
        "Windows requires the Scale plan or a paid add-on"
      ],
      "verdict": "A persistent-desktop service for computer-use agents with published per-second pricing, suited to agents that need an always-on computer; each computer is described as an isolated VM and the hypervisor is not named in the docs.",
      "peers": [
        "scrapybara",
        "cua",
        "exe-dev",
        "kasm-workspaces",
        "kernel"
      ],
      "sources": [
        "https://www.orgo.ai",
        "https://www.orgo.ai/pricing.md",
        "https://www.orgo.ai/llms.txt",
        "https://docs.orgo.ai",
        "https://docs.orgo.ai/llms.txt",
        "https://docs.orgo.ai/llms-full.txt",
        "https://trust.orgo.ai/",
        "https://docs.orgo.ai/guides/browser-networking"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "kasm-workspaces",
      "name": "Kasm Workspaces",
      "company": "Kasm Technologies",
      "hq": {
        "city": "McLean",
        "region": "Virginia",
        "country": "US",
        "source": "https://kasm.com/contact"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://kasm.com",
      "docs": "https://docs.kasm.com",
      "pricing_url": "https://kasm.com/community-edition",
      "github": "https://github.com/kasmtech",
      "tagline": "Container-streamed Linux desktops and browsers, self-hosted or as Kasm SaaS, with a Developer API and a free Community Edition that Kasm also markets for agentic AI.",
      "isolation": "Docker container per session, streamed to the browser with KasmVNC; sessions can also be RDP-gatewayed to Windows hosts",
      "isolation_short": "Docker container",
      "isolation_class": "container",
      "open_source": false,
      "license": null,
      "self_host": true,
      "byoc": null,
      "sdks": [
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Session expiry is governed by per-user usage limits; the `keepalive` API resets a session's expiration time",
      "max_session_short": "Admin-defined",
      "persistence": "Sessions are ephemeral containers; persistent profiles can be enabled per image so user data survives across sessions. Session recording available on paid tiers",
      "snapshots": null,
      "pause_resume": null,
      "gpu": true,
      "gpu_detail": "GPU resources can be shared across users and workloads; release 1.19 adds SR-IOV bifurcation of Intel Arc Pro cards (press release)",
      "networking": "Administrators control inbound and outbound access per workspace; web filtering; optional PureVPN managed egress on Essentials and Premier",
      "pricing": "Community Edition free (up to 5 concurrent sessions, non-commercial). Basic, Essentials and Premier editions: \"Contact Us\", licensed per named user or per concurrent session; Cloud and Enterprise SaaS available",
      "free_tier": "Community Edition: free for individuals, nonprofits and non-commercial use, up to 5 concurrent sessions",
      "compliance": null,
      "regions": null,
      "best_for": "Organisations that already run browser-streamed desktops and want to give agents the same isolated, policy-controlled containers",
      "overview": [
        "Kasm Workspaces, from Kasm Technologies (McLean, Virginia), is a \"container streaming platform\": each session is a Docker container running a desktop, browser or single application, rendered to the user's browser through KasmVNC so nothing runs on the endpoint. It installs on a single Linux server or on Kubernetes, can autoscale into cloud providers, and is also sold as Kasm-hosted Cloud and Enterprise SaaS. A Developer API (`request_kasm`, `get_kasm_status`, `get_kasm_screenshot`, `exec_command_kasm`, `destroy_kasm`, `keepalive`) lets integrators create sessions for a user, run commands, take JPEG screenshots and tear sessions down programmatically.",
        "Kasm's \"Secure Agentic AI\" page describes computer-use agents and workspace-native agents each running in its own ephemeral container, with credentials held inside the workspace and administrators controlling network egress; it does not document a purpose-built agent SDK or named model integrations. Source for the core server is not published, while workspace images, KasmVNC, Ansible and Terraform automations are on GitHub. The Community Edition is free for non-commercial use up to 5 concurrent sessions; commercial editions are quote-based and licensed per named user or per concurrent session."
      ],
      "notable": [
        "Community Edition is free for individuals and nonprofits with a 5-concurrent-session cap; commercial editions are priced on request",
        "Current release is 1.19.0 with rolling images by default (install docs)",
        "Workspace images (1.2k stars) and KasmVNC (5.4k stars) are open on GitHub; source for the core platform is not published there"
      ],
      "pros": [
        "You want self-hosted, browser-streamed desktop containers with admin policy controls and a session API",
        "You need a free, full-featured edition for testing or non-commercial agent experiments",
        "You run on Kubernetes or want Terraform/Ansible-driven deployment with cloud autoscaling"
      ],
      "cons": [
        "An agent-specific SDK or named model integrations are not documented; agents use the general Developer API",
        "Container sessions use Docker (Sysbox runtime optional), with VM-backed server workspaces as a separate type; startup figures are not documented",
        "Commercial pricing is not published; the Community Edition is capped at 5 concurrent sessions and non-commercial use"
      ],
      "verdict": "A container-streaming workspace platform whose Developer API can hand agents isolated desktops, best suited to teams that already operate it.",
      "peers": [
        "cua",
        "orgo",
        "scrapybara",
        "hyperbeam",
        "browserless",
        "coder"
      ],
      "sources": [
        "https://kasm.com/",
        "https://kasm.com/community-edition",
        "https://kasm.com/solutions/ai-environments/secure-agentic-ai",
        "https://docs.kasm.com/docs/reference/license",
        "https://docs.kasm.com/docs/reference/developer-api",
        "https://docs.kasm.com/docs/tutorials/install/single-server-install",
        "https://github.com/kasmtech",
        "https://docs.kasm.com/docs/how-to/workspaces-sessions/sessions/",
        "https://docs.kasm.com/docs/how-to/infra-autoscale/servers",
        "https://docs.kasm.com/docs/how-to/workspaces-sessions/container-workspace/sysbox-runtime",
        "https://kasm.com/press/kasm-intel-partnership-private-ai-xeon-6-amx"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "hyperbeam",
      "name": "Hyperbeam",
      "company": "Hyperbeam Inc.",
      "hq": {
        "city": "Waterloo",
        "region": "Ontario",
        "country": "CA",
        "source": "https://www.ycombinator.com/companies/hyperbeam"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://hyperbeam.com",
      "docs": "https://docs.hyperbeam.com",
      "pricing_url": "https://hyperbeam.com/#pricing",
      "github": null,
      "tagline": "Embeddable hosted virtual computers (Chromium browsers, plus Android and console emulators on request) created by REST API and streamed into web apps for multi-user control.",
      "isolation": "Not documented; the vendor calls each session a \"virtual computer\" and the REST endpoint is /vm, but no hypervisor or container runtime is named",
      "isolation_short": "Not documented",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "REST API",
        "JavaScript/TypeScript",
        "Unity"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "No absolute limit by default: `timeout.absolute` and `timeout.inactive` default to null; `timeout.offline` (no connected users) defaults to 3,600 s",
      "max_session_short": "Unlimited by default",
      "persistence": "Optional Chrome profile persistence (bookmarks, history, cookies, passwords) stored encrypted in S3 and reloaded by session ID; unused profiles expire after three months",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Home page mentions \"custom-built servers made for GPU-intensive workloads\" for WebGL; `webgl` is a per-session flag",
      "networking": "Outbound internet from the hosted browser; regions NA, EU and AS selectable per session; a PAC-based `http_proxy` option is described in the FAQ; domain filtering is not documented",
      "pricing": "Pay as you go after a free allowance; the home-page estimator shows $0.007 per participant-minute with automatic volume discounts and custom pricing at high volume",
      "free_tier": "10,000 free minutes every month, no credit card required (home page)",
      "compliance": null,
      "regions": "NA (default), EU and AS",
      "best_for": "Products that embed a shared, streamed browser for users, where an agent may also drive it through the client SDK",
      "overview": [
        "Hyperbeam provides \"virtual computers\" that developers embed in their own websites and apps. A `POST` to the REST API starts a Chromium browser session (Android and NES emulators are available on request) and returns an `embed_url` for the `@hyperbeam/web` client, plus an `admin_token` that unlocks programmatic navigation and permission control. Sessions are tuned at creation with resolution (up to 1920x1080), frame rate (24 to 60), kiosk mode, extensions, locale, search engine and region (NA, EU or AS). Hyperbeam also runs a consumer Watch Party platform, and the API is built around multi-user streaming with synchronised audio and video.",
        "Timeouts are configurable per session: an absolute cap, an inactivity cap and an offline cap (default one hour with no connected users), with webhooks for timeout warnings. Chrome profiles can be persisted to encrypted S3 storage and reloaded by session ID. The home page advertises 10,000 free minutes a month and pay-as-you-go at $0.007 per minute thereafter. Hyperbeam's home page and docs do not mention AI agents, and a CDP endpoint, Playwright connection or server-side screenshot API is not documented; the client SDK's `sendEvent` method sends keyboard, mouse and wheel events and `frameCb` exposes video frames. The isolation technology is not described."
      ],
      "notable": [
        "10,000 free minutes per month, then $0.007 per participant-minute with automatic volume discounts (home page)",
        "Sessions default to no absolute time limit; the offline timeout defaults to 3,600 seconds",
        "Persisted browser profiles are deleted after three months without access"
      ],
      "pros": [
        "You want to embed a streamed browser in your own product that humans and an agent can share",
        "You need EU or Asia hosting and fine control over resolution, frame rate and timeouts",
        "You are prototyping and want a large monthly free allowance"
      ],
      "cons": [
        "AI-agent use, CDP and Playwright access are not documented; input events go through the client-side SDK",
        "Isolation technology and compliance attestations are not documented",
        "Browser only by default; Android and console emulators require access requests"
      ],
      "verdict": "An API for embedding multi-user, streamed virtual browsers in web apps, suited to products that need shared browsing; AI-agent use and CDP or Playwright access are not documented.",
      "peers": [
        "browserbase",
        "steel",
        "browserless",
        "hyperbrowser",
        "anchor-browser",
        "kasm-workspaces"
      ],
      "sources": [
        "https://hyperbeam.com",
        "https://docs.hyperbeam.com",
        "https://docs.hyperbeam.com/home/getting-started",
        "https://docs.hyperbeam.com/rest-api/dispatch/start-chromium-session",
        "https://docs.hyperbeam.com/guides/timeouts",
        "https://docs.hyperbeam.com/guides/persist-session-state",
        "https://docs.hyperbeam.com/home/faq",
        "https://docs.hyperbeam.com/client-sdk/javascript/reference",
        "https://docs.hyperbeam.com/rest-api/dispatch/start-android-emulator-session",
        "https://docs.hyperbeam.com/rest-api/dispatch/start-console-emulator-session"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "judge0",
      "name": "Judge0",
      "company": "Judge0 d.o.o.",
      "hq": {
        "city": "Zagreb",
        "country": "HR",
        "source": "https://judge0.com"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://judge0.com",
      "docs": "https://ce.judge0.com",
      "pricing_url": "https://judge0.com/#pricing",
      "github": "https://github.com/judge0/judge0",
      "tagline": "Open-source online code execution system (GPL-3.0) that runs each submission in an isolate sandbox, offered as Judge0 Cloud via RapidAPI or directly and as a self-hosted Docker Compose stack.",
      "isolation": "isolate sandbox (Linux namespaces and control groups) per submission; Judge0 workers run as Docker Compose services and compilation happens inside the sandbox since v1.4.0",
      "isolation_short": "isolate (namespaces/cgroups)",
      "isolation_class": "os",
      "open_source": true,
      "license": "GPL-3.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "REST API",
        "Python",
        "JavaScript/TypeScript"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Per submission: cpu_time_limit default 2 s (max 15 s) and wall_time_limit default 5 s (max 20 s) on the documented defaults; administrators of self-hosted instances can change the maxima",
      "max_session_short": "5 s wall (max 20 s)",
      "persistence": "None; each submission is an independent run identified by a token, with inputs supplied as source_code, stdin and an optional Base64 zip of additional_files",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": "No GPU option is documented",
      "networking": "enable_network defaults to true on the documented configuration; administrators can disallow it (ALLOW_ENABLE_NETWORK / ENABLE_NETWORK, added in v1.13.0)",
      "pricing": "Direct Judge0 Cloud plans (judge0.com, EUR, monthly): Pro €27 for 2,000 submissions/day, Ultra €54 for 5,000/day, Mega €107 for 10,000/day, €0.001 per extra submission. RapidAPI listing for Judge0 CE: Basic $0, Pro $44.99, Ultra $89.99, Mega $169.99 per month (daily caps not retrievable from the listing without JavaScript). Self-hosting is free.",
      "free_tier": "RapidAPI Basic plan at $0/month (the CE docs say to \"start with the FREE Basic Plan\"); the official judge0/examples README also points to Sulu with \"20K Free Submissions\"; free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Batch \"submit code, get stdout\" judging in 60-90+ languages with per-submission CPU, wall-clock and memory limits",
      "overview": [
        "Judge0 is an open-source online code execution system created by Herman Zvonimir Došilović (established August 2016) and run by Judge0 d.o.o. in Zagreb, Croatia. It is an HTTP JSON API: you POST a submission with `language_id`, `source_code` and optional `stdin`, receive a token, and poll for stdout, stderr, exit code, CPU time and memory, or supply a `callback_url` to be notified. It comes in two flavours, Judge0 CE and Judge0 Extra CE, which differ in the supported languages (the homepage says 60+, the README 90+). Each submission is compiled and run inside an [isolate](https://github.com/ioi/isolate) sandbox, which the isolate project describes as using Linux namespaces and control groups; the service itself is deployed as Docker Compose services (API, workers, PostgreSQL, Redis) and the changelog says it has only been tested on Linux.",
        "There is no state between calls: every submission is a fresh run bounded by `cpu_time_limit` (default 2 s, maximum 15 s), `wall_time_limit` (default 5 s, maximum 20 s), `memory_limit` (default 128,000 KB, maximum 256,000 KB), `max_file_size` and `max_processes_and_or_threads`, all of which an administrator can raise on a self-hosted instance. Network access is governed by the `enable_network` flag added in v1.13.0. The hosted service is sold through RapidAPI (CE and Extra CE listings, Basic plan at $0) and directly from judge0.com in EUR tiers; the GitHub README positions it \"for humans and AI\" and lists AI agents first among use cases. The code is GPL-3.0; official Python (`pip install judge0`, MIT) and JavaScript SDKs cover RapidAPI, AllThingsDev and self-hosted endpoints. The latest tagged release is v1.13.1 (18 April 2024), which fixed three CVEs; the repository was still receiving commits in October 2026."
      ],
      "notable": [
        "Established August 2016; the design is described in the MIPRO 2020 paper \"Robust and Scalable Online Code Execution System\"",
        "Latest release v1.13.1 (18 April 2024) fixed CVE-2024-28185, CVE-2024-28189 and CVE-2024-29021; versions before 1.13.0 receive no security support",
        "About 4.5k GitHub stars; the README cites dozens of academic papers that use Judge0"
      ],
      "pros": [
        "You want a mature, widely integrated submit-and-poll judge with explicit CPU, wall-clock, memory and file-size limits per run",
        "You need 60-90+ languages, including compiled ones, behind one JSON API",
        "You run on-premise and want a GPL-3.0 stack you can deploy with Docker Compose",
        "You are building batch evaluation and want batched submissions and webhooks rather than long-lived sessions"
      ],
      "cons": [
        "Stateless by design: no filesystem or process survives between submissions, so multi-step agent work must re-send everything each time",
        "Documented defaults cap a run at 5 s wall time (20 s maximum) and 256 MB on the hosted configuration",
        "isolate is a process-level sandbox (namespaces and cgroups), not a VM; the self-host guide for Ubuntu 22.04 requires disabling the unified cgroup v2 hierarchy",
        "GPL-3.0 licence and the last tagged release dating from April 2024; no compliance certifications are published"
      ],
      "verdict": "The reference open-source online judge: ideal for stateless, time-boxed code evaluation at scale, but not a workspace an agent can return to.",
      "peers": [
        "piston",
        "sphere-engine",
        "onecompiler",
        "jdoodle",
        "codapi",
        "e2b"
      ],
      "sources": [
        "https://judge0.com",
        "https://github.com/judge0/judge0",
        "https://ce.judge0.com",
        "https://github.com/judge0/judge0/blob/master/CHANGELOG.md",
        "https://github.com/judge0/judge0/blob/master/SECURITY.md",
        "https://github.com/judge0/judge0-python",
        "https://github.com/judge0/examples",
        "https://github.com/ioi/isolate",
        "https://rapidapi.com/judge0-official/api/judge0-ce/pricing",
        "https://github.com/judge0/judge0/releases"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "piston",
      "name": "Piston",
      "company": "Engineer Man",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/engineer-man/piston",
      "docs": "https://github.com/engineer-man/piston#readme",
      "pricing_url": null,
      "github": "https://github.com/engineer-man/piston",
      "tagline": "MIT-licensed code execution engine that runs each submission in an isolate sandbox inside a Docker container, with a public API at emkc.org restricted to approved non-commercial projects since February 2026.",
      "isolation": "isolate inside Docker: Linux namespaces, chroot, a different unprivileged user per submission and cgroups for resource limits; network disabled by default; the container itself runs with --privileged",
      "isolation_short": "isolate in Docker",
      "isolation_class": "os",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Per request: run_timeout default 3,000 ms and compile_timeout default 10,000 ms wall time, with matching run_cpu_time / compile_cpu_time; memory limits default to -1 (no limit)",
      "max_session_short": "3 s run / 10 s compile",
      "persistence": "None; the README says submissions \"shall not be aware of each other\" and temp space is cleaned after each run. A WebSocket endpoint (/api/v2/connect) lets you send stdin to a process while it runs",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": "No GPU option is documented",
      "networking": "Outgoing network disabled by default",
      "pricing": "Free to self-host (MIT). The public emkc.org API has no listed price; keys are granted at the maintainer's discretion only for non-commercial educational projects",
      "free_tier": "Free to self-host; public API keys are not freely available as of 15 February 2026",
      "compliance": null,
      "regions": null,
      "best_for": "Self-hosted, stateless execution of untrusted snippets in 80+ languages with isolate-level sandboxing",
      "overview": [
        "Piston is a general-purpose code execution engine from the Engineer Man community, used for EMKC coding challenges, a Discord bot and, per the README, 200+ direct integrations. You run it as the `ghcr.io/engineer-man/piston` Docker image (started with `--privileged`), install language runtimes with the `ppman` CLI, then `POST /api/v2/execute` with a `language`, `version` and a `files` array; the response contains `compile` and `run` objects with stdout, stderr, exit code, signal and a status such as `TO` for timeout. The README lists 84 language entries.",
        "Sandboxing is [isolate](https://www.ucw.cz/moe/isolate.1.html) inside Docker: each submission gets its own Linux namespaces, a chroot, a distinct unprivileged user, a cgroup-enforced memory cap, a process limit (256), an open-file limit (2,048), stdout capped at 1,024 characters by default and outgoing network disabled. Runs are wall-clock bounded (3 s run, 10 s compile by default) and nothing persists afterwards. Self-hosting needs cgroup v2 enabled and v1 disabled. The README says the public API at emkc.org is \"no longer freely available to the public (as of Feb 15, 2026)\"; authorisation is granted only for \"good cause non-commercial educational projects\", and keys are not granted for \"Projects that cost money\"."
      ],
      "notable": [
        "Public API keys no longer freely available as of 15 February 2026; granted only for non-commercial educational projects",
        "About 2.8k GitHub stars; the repository was still receiving commits in July 2026",
        "README lists 84 supported languages installed on demand through ppman"
      ],
      "pros": [
        "You want a permissively licensed (MIT) execution engine you can run yourself with one Docker command",
        "You need isolate-grade process sandboxing with network off and per-submission unprivileged users",
        "You are evaluating many short snippets across dozens of languages and compiled toolchains"
      ],
      "cons": [
        "Public emkc.org API needs an authorisation key since 15 February 2026, granted only for non-commercial educational projects",
        "Stateless: no files, packages or processes survive between requests",
        "Defaults are 3 s per run and 1,024 characters of stdout; the documented docker run command uses --privileged",
        "Client libraries are third-party wrappers that the README lists as approved extensions; the core interface is the HTTP API plus a WebSocket for interactive stdin (self-hosted only)"
      ],
      "verdict": "A self-hosted engine for stateless snippet execution; the public emkc.org API requires an authorisation key that the README says is granted only for non-commercial educational projects.",
      "peers": [
        "judge0",
        "codapi",
        "glot",
        "onecompiler",
        "jdoodle"
      ],
      "sources": [
        "https://github.com/engineer-man/piston",
        "https://raw.githubusercontent.com/engineer-man/piston/master/readme.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "sphere-engine",
      "name": "Sphere Engine",
      "company": "Sphere Research Labs",
      "hq": {
        "city": "Gdynia",
        "country": "PL",
        "source": "https://sphere-engine.com/contact"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://sphere-engine.com",
      "docs": "https://docs.sphere-engine.com",
      "pricing_url": "https://sphere-engine.com/pricing",
      "github": null,
      "tagline": "Commercial code execution service from the makers of Ideone and SPOJ: a Compilers API for 80+ languages, a Problems API for test-case grading and a Containers product for multi-file projects, priced by quote.",
      "isolation": "Vendor does not name the technology; the FAQ says each submission runs in \"an isolated workspace dedicated to a single submission\" with exclusive access to one physical CPU core, and the Containers product runs multi-file projects in configurable slots",
      "isolation_short": "Isolated workspace (unnamed)",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": true,
      "byoc": null,
      "sdks": [
        "REST API",
        "PHP",
        "Python",
        "JavaScript/TypeScript"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Compilers: timeLimit default 5 s, configurable between 1 and 15 s; Problems: 1-60 s per test case; Containers: timeout set per execution stage",
      "max_session_short": "5 s (1-15 s)",
      "persistence": "Compilers and Problems submissions are independent runs whose source and streams are retained for one month; Containers workspaces can be stopped and resumed through the API or panel",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "No GPU option is documented",
      "networking": "Compilers and Problems: no Internet access by default, can be enabled for subscribed users; Containers: network is a per-project setting",
      "pricing": "Individual quotes based on project scope, number of students or candidates, expected API submissions and support level; annual and quarterly plans; cloud or on-premise deployment",
      "free_tier": "14-day fully functional free trial, no credit card required",
      "compliance": null,
      "regions": null,
      "best_for": "Assessment and e-learning products that need a supported, commercially backed compile-and-run API with test-case grading",
      "overview": [
        "Sphere Engine is the commercial code-execution platform of Sphere Research Labs Sp. z o.o. (Gdynia, Poland), the company behind the free Ideone and Sphere Online Judge services. The Compilers API (v4) is asynchronous: `POST /submissions` with `compilerId`, source and optional input to a customer-specific endpoint returns an id, then you poll `GET /submissions/{id}` every couple of seconds or receive a webhook; results include status codes, time, memory, signal and output streams. The Problems API adds test cases and grading, and the newer Containers product runs multi-file projects with an embeddable workspace (editor plus Linux terminal). Official client libraries exist for PHP and Python (the Java one is marked suspended) alongside JavaScript widget SDKs.",
        "Submissions are stateless runs with a default `timeLimit` of 5 s (1-15 s on Compilers, up to 60 s per Problems test case), 1.5 GB RAM, 64 kB source and 64 kB input; Internet access is off unless enabled for a subscriber. The FAQ describes an isolated per-submission workspace with a dedicated physical core but does not name the sandboxing technology. Pricing is by quote after a 14-day trial, with on-premise deployment available; the homepage claims 5,000,000+ submissions per month. Agent-specific positioning does not appear on the homepage or the documentation pages checked."
      ],
      "notable": [
        "Homepage claims 5,000,000+ code submissions per month and 2,000,000+ end users",
        "Compilers API v4 is not backward compatible with v3; official PHP and Python clients, Java client suspended",
        "Submission data (source and streams) retained for one month by default"
      ],
      "pros": [
        "You want a vendor-supported compile-and-run API with SLA, on-premise option and test-case grading",
        "You need 80+ languages and a webhook-driven asynchronous flow",
        "You are building assessments or courses and want the embeddable editor and Containers workspaces too"
      ],
      "cons": [
        "Pricing is quote-only and the sandbox technology is not documented",
        "Compilers runs are capped at 15 s and 1.5 GB with no Internet by default",
        "No persistent state between Compilers submissions; only Containers workspaces can be stopped and resumed",
        "Agent-framework integrations and an MCP server are not documented"
      ],
      "verdict": "A commercially supported judge for assessment platforms; agents get a reliable but stateless and quote-priced execution backend.",
      "peers": [
        "judge0",
        "onecompiler",
        "jdoodle",
        "piston",
        "codesandbox"
      ],
      "sources": [
        "https://sphere-engine.com",
        "https://sphere-engine.com/pricing",
        "https://docs.sphere-engine.com/compilers/overview",
        "https://docs.sphere-engine.com/compilers/api/quickstart",
        "https://docs.sphere-engine.com/compilers/api/overview-version-4",
        "https://docs.sphere-engine.com/sphere-engine-faq",
        "https://docs.sphere-engine.com/containers/overview",
        "https://ideone.com"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "onecompiler",
      "name": "OneCompiler Code Execution API",
      "company": "OneCompiler",
      "hq": {
        "city": "Hyderabad",
        "region": "Telangana",
        "country": "IN",
        "source": "https://onecompiler.com/about"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://onecompiler.com/apis",
      "docs": "https://onecompiler.com/apis/code-execution",
      "pricing_url": "https://onecompiler.com/apis/pricing",
      "github": null,
      "tagline": "Hosted run-code API behind the OneCompiler online IDE: one POST runs a multi-file program in 150+ languages and databases, billed per call or per credit with enterprise tiers that state a maximum runtime of 7-20 s.",
      "isolation": "Not documented; the API pages describe no sandbox or isolation mechanism",
      "isolation_short": "Not documented",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Standard plans: not stated (a timeout error code E001 exists). Enterprise usage plans state a max runtime of 7 s (E1, E2), 10 s (E3) or 20 s (E4)",
      "max_session_short": "7-20 s (enterprise)",
      "persistence": "None; each call runs the submitted files once per stdin value. A `persist` flag keeps the output URL for 1 hour",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": "No GPU option is documented",
      "networking": "Not documented for Code Execution API runs; the pricing page lists \"Internet Permission (In Editor)\" as included only on the E4 and EU3 tiers and their custom plans",
      "pricing": "Code Execution plans: P1 $5/month for 10,000 calls, P2 $10 for 30,000, P3 $25 for 100,000, P4 $50 for 300,000. Credit model on the API page: $1 for 2,000 credits ($100+ buys 4,000 per $1, $500+ buys 6,000 per $1). Enterprise usage plans E1 $100 (500,000 calls) to E4 $2,000 (10,000,000 calls) per month; user-based plans $2-12 per user per year. Also sold via RapidAPI and Zyla API Hub",
      "free_tier": "100 free credits for new accounts",
      "compliance": null,
      "regions": null,
      "best_for": "Cheap, high-volume stateless runs across 150+ languages and databases from a backend",
      "overview": [
        "OneCompiler is an online IDE whose Code Execution API exposes the same runtimes to backend applications: `POST https://api.onecompiler.com/v1/run` with an `X-API-Key` header, a `language`, a `files` array of name/content objects and optional `stdin` (a string or an array, in which case the program runs once per value). The response reports stdout, stderr, compilation and execution time in milliseconds and memory used. The API covers 150+ languages plus web languages and databases, and the API Console manages keys, credits and usage.",
        "Runs are stateless and quotas are enforced per plan (error E002 \"API quota exceeded\"); the documentation does not describe the sandbox, memory cap or default timeout, although enterprise tiers specify maximum runtimes of 7, 10 or 20 s and the pricing page lists \"Internet Permission (In Editor)\" only on the E4, EU3 and custom plans. Pricing is unusually explicit for the category: $5 per month for 10,000 calls up to $50 for 300,000, a credit model at $1 per 2,000 credits, and enterprise tiers up to $2,000 per month for 10 million calls. STDIN is accepted up to 1,000,000 characters (10,000 on marketplace listings). SDKs, a self-hosting option and compliance information are not documented."
      ],
      "notable": [
        "Published call-based pricing from $5/month (10,000 calls) to $2,000/month (10,000,000 calls)",
        "Enterprise tiers list a maximum runtime of 7 s (E1/E2), 10 s (E3) and 20 s (E4), and \"Internet Permission (In Editor)\" from E4",
        "STDIN accepted up to 1,000,000 characters, with batch runs by passing an array of inputs"
      ],
      "pros": [
        "You want the lowest published per-call price for stateless runs",
        "You need 150+ languages and databases (SQL engines included) through one endpoint",
        "You are batch-testing one program against many inputs in a single request"
      ],
      "cons": [
        "No documentation of the sandbox, memory limit or default timeout",
        "\"Internet Permission (In Editor)\" is listed only on the E4, EU3 and custom plans; Internet access for API runs is not documented",
        "Stateless, with no sessions, files or packages persisting between calls",
        "No SDKs, self-hosting or compliance attestations are published"
      ],
      "verdict": "A low-priced run-code endpoint with published per-call pricing; the sandbox is not documented, and it suits simple stateless tool calls rather than workspaces.",
      "peers": [
        "jdoodle",
        "judge0",
        "sphere-engine",
        "piston",
        "codapi"
      ],
      "sources": [
        "https://onecompiler.com/apis",
        "https://onecompiler.com/apis/code-execution",
        "https://onecompiler.com/apis/pricing",
        "https://onecompiler.com"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "jdoodle",
      "name": "JDoodle Compiler API",
      "company": "JDoodle Labs Pty Ltd",
      "hq": {
        "region": "Australian Capital Territory",
        "country": "AU",
        "source": "https://abr.business.gov.au/ABN/View?abn=58700156725"
      },
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://www.jdoodle.com/compiler-api",
      "docs": "https://www.jdoodle.com/docs/api",
      "pricing_url": "https://www.jdoodle.com/pricing",
      "github": null,
      "tagline": "Credit-metered compile-and-run REST and WebSocket API from the JDoodle online IDE, covering 114 languages, frameworks and databases, with each run costing one credit and a free daily allowance.",
      "isolation": "Vendor does not name the technology; the product page says each run gets \"a fresh sandbox with its own filesystem and process limits\" that can stop a never-ending program and has no route to your network",
      "isolation_short": "Fresh sandbox per run (unnamed)",
      "isolation_class": null,
      "open_source": false,
      "license": null,
      "self_host": false,
      "byoc": null,
      "sdks": [
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "None; each REST call is one run. A WebSocket mode streams output and accepts input while a program runs, and multi-file projects can be uploaded for one credit extra",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": "No GPU option is documented",
      "networking": "Internet access is available on paid API plans only and costs one extra credit per run that uses it",
      "pricing": "One credit per run; opening a WebSocket connection, uploading a multi-file project and enabling Internet access each add one credit; credits reset daily. Paid API tiers are a separate subscription from the platform plans: Starter $10/month for 1,000 credits per day, Growth $20 for 2,000, Scale $50 for 5,000, Business $100 for 10,000, Volume $200-1,000 for 20,000-100,000 and Custom by quote",
      "free_tier": "Free API plan with 20 credits per day; Internet access is not available on the Free plan",
      "compliance": null,
      "regions": null,
      "best_for": "Low-volume, pay-per-run execution with an interactive stdin stream for teaching and demo products",
      "overview": [
        "JDoodle Labs Pty Ltd (Australia) runs the JDoodle online IDE and sells its Compiler API separately. `POST https://api.jdoodle.com/v1/execute` returns output, status code, memory and CPU time in one response, and a WebSocket connection streams output as it is produced and lets the user answer prompts. The API covers 114 languages, frameworks and databases with multiple selectable versions. The product page describes a fresh sandbox per run with its own filesystem and process limits, which can kill non-terminating programs and has no route to your network; the technology is not named.",
        "Billing is a daily-resetting credit model: one run is one credit, and a WebSocket session, a multi-file upload or Internet access each add one more. The free API plan grants credits every day; paid tiers are a separate monthly or yearly subscription from the platform plans (Free, Student $4, Pro $12, Teacher $20 per month). Execution time and memory limits are not published. There is no SDK, self-hosting or compliance documentation, and no state persists between calls."
      ],
      "notable": [
        "Credits reset daily; one run equals one credit, with WebSocket, multi-file and Internet access each costing one extra credit",
        "114 languages, frameworks and databases, each with several versions",
        "Platform plans from US$0 (1,000 code executions per day) to Teacher US$20/month; the API is billed separately"
      ],
      "pros": [
        "You want a simple per-run API with a free daily allowance for prototyping",
        "You need interactive programs: the WebSocket mode streams output and accepts input mid-run",
        "You are embedding execution in an education product alongside the embeddable IDE"
      ],
      "cons": [
        "Execution time, memory and output limits are not documented",
        "Sandbox technology is not named; isolation is described only as a fresh filesystem and process limits",
        "Stateless; Internet access costs an extra credit per run and uploading a multi-file project costs one credit",
        "SDKs and compliance information are not documented"
      ],
      "verdict": "A credit-metered compiler API for demos and courses; execution time, memory and output limits and the sandbox technology are not documented.",
      "peers": [
        "onecompiler",
        "judge0",
        "sphere-engine",
        "piston",
        "codapi"
      ],
      "sources": [
        "https://www.jdoodle.com/compiler-api",
        "https://www.jdoodle.com/docs/api",
        "https://www.jdoodle.com/pricing",
        "https://www.jdoodle.com/docs/getting-started/plans-billing/getting-started/",
        "https://www.jdoodle.com/docs/api/credits",
        "https://www.jdoodle.com/docs/api/errors",
        "https://www.jdoodle.com/docs/api/rest",
        "https://www.jdoodle.com/docs/api/websocket"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "codapi",
      "name": "Codapi",
      "company": "Codapi (Anton Zhiyanov)",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://codapi.org",
      "docs": "https://codapi.org/docs/",
      "pricing_url": "https://codapi.org/pricing/",
      "github": "https://github.com/nalgeon/codapi",
      "tagline": "Lightweight Apache-2.0 sandbox server (a 3 MB Go binary) that runs each snippet in a locked-down Docker container, built for interactive documentation, with a hosted cloud at $5/month and free self-hosting.",
      "isolation": "Docker (or Podman/OrbStack) container per run with the shipped defaults: runc runtime, 1 CPU, 64 MB memory, network none, read-only root with the sandbox volume mounted read-only, all capabilities dropped, nofile 96, nproc 64, non-root \"sandbox\" user, 5 s timeout, 4,096-byte output cap",
      "isolation_short": "Locked-down Docker container",
      "isolation_class": "container",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "REST API",
        "JavaScript/TypeScript"
      ],
      "cold_start": "FAQ: server-side sandboxes take 1-3 seconds per run depending on hardware; in-browser playgrounds run almost instantly",
      "cold_start_short": "1-3 s per run (claim)",
      "max_session": "Per step: default timeout 5 s in the shipped codapi.json, configurable per sandbox",
      "max_session_short": "5 s default",
      "persistence": "None; each /v1/exec call runs the submitted files once in a fresh container and returns id, ok, duration, stdout and stderr",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": "No GPU option is documented",
      "networking": "Network \"none\" in the default box configuration; sandboxes such as curl, Fetch and HTTP are provided for network-related examples",
      "pricing": "Cloud $5/month for 5,000 runs/month and 30+ sandboxes; self-hosted free and unlimited; Codapi+ commercial support $250/month. The pricing page notes pricing \"may (and probably will) change\"",
      "free_tier": "Cloud: free to try without sign-up, 1,000 code runs per month; free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Embedding runnable snippets in docs and courses, or a tiny self-hosted run-code endpoint with strict defaults",
      "overview": [
        "Codapi is an open-source sandbox server by Anton Zhiyanov for interactive code examples in documentation, courses and blog posts. The server is a 3 MB Go binary that manages sandboxes, each one or more Docker images registered via `box.json` and `commands.json`; the companion 5 KB `codapi-js` widget turns static code blocks into runnable ones. The API is a single `POST /v1/exec` with `sandbox`, `command` and a `files` map, returning `ok`, `duration`, `stdout` and `stderr`. About 40 playgrounds ship out of the box, covering mainstream and emerging languages, databases such as PostgreSQL, DuckDB and ClickHouse, and tools like Git and LaTeX; custom sandboxes are any Dockerfile.",
        "The shipped `codapi.json` is the most explicit default hardening in this category: 1 CPU, 64 MB RAM, network none, read-only filesystem, all capabilities dropped, 64 processes, 96 open files, non-root user, 5 s timeout and 4 KB of output per step. The FAQ says server-side runs take 1-3 seconds and that the sandbox server should not have access to sensitive data. Runs are stateless. The cloud service is free to try (1,000 runs per month) and $5/month for 5,000 runs; self-hosting is free and unlimited, with docs for Debian, Nginx, XFS quotas and running locally with Docker, Podman, OrbStack or Lima on a Mac. Latest release v0.14.2 (22 August 2026); about 2.1k GitHub stars."
      ],
      "notable": [
        "Latest release v0.14.2 on 22 August 2026; about 2.1k GitHub stars",
        "Default per-run limits in the shipped config: 1 CPU, 64 MB, no network, 5 s timeout, 4,096-byte output",
        "Cloud plan $5/month for 5,000 runs; the pricing page warns prices will probably change"
      ],
      "pros": [
        "You want a tiny, auditable self-hosted execution endpoint with hardened container defaults",
        "You need runnable examples embedded in documentation or a course",
        "You are running databases and CLI tools (PostgreSQL, DuckDB, Git) as sandboxes, not only languages"
      ],
      "cons": [
        "Stateless by design: no sessions, and the defaults allow 64 MB and 5 s per run",
        "Default isolation is a restricted Docker container on the runc runtime, not a VM or user-space kernel",
        "The paid cloud plan covers 5,000 runs per month, and the pricing page says Codapi is still in its early stages",
        "A Python or other backend SDK is not documented; the interfaces are the HTTP API and the JavaScript widget"
      ],
      "verdict": "The simplest well-hardened run-code server to self-host; fine for doc snippets and stateless tool calls, not for agent workspaces.",
      "peers": [
        "piston",
        "judge0",
        "glot",
        "onecompiler",
        "docker-sandboxes"
      ],
      "sources": [
        "https://codapi.org",
        "https://codapi.org/pricing/",
        "https://codapi.org/faq/",
        "https://codapi.org/docs/",
        "https://github.com/nalgeon/codapi",
        "https://raw.githubusercontent.com/nalgeon/codapi/main/codapi.json",
        "https://raw.githubusercontent.com/nalgeon/codapi/main/docs/api.md",
        "https://raw.githubusercontent.com/nalgeon/codapi/main/docs/add-sandbox.md",
        "https://raw.githubusercontent.com/nalgeon/codapi/main/docs/production.md",
        "https://github.com/nalgeon/codapi/blob/main/LICENSE"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "glot",
      "name": "glot.io",
      "company": "glot.io (Petter Rasmussen)",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://glot.io",
      "docs": "https://github.com/glotcode/docker-run#readme",
      "pricing_url": null,
      "github": "https://github.com/glotcode/glot",
      "tagline": "Open-source code playground whose runner, docker-run (MIT), executes every request in a fresh, single-use Docker container with network disabled; 44 languages on the hosted site.",
      "isolation": "Fresh single-use Docker container per run, deleted afterwards; docker-run configures network off, a memory cap, nproc and nofile ulimits, dropped capabilities and an optional read-only root filesystem; the hosted site says code \"runs in a temporary container with network access disabled\"",
      "isolation_short": "Single-use Docker container",
      "isolation_class": "container",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Per run: RUN_MAX_EXECUTION_TIME (seconds) set by the operator; the container is killed when it is reached. No figure is published for the hosted site",
      "max_session_short": "Operator-set per run",
      "persistence": "None; containers are never reused, and prewarmed containers serve exactly one request. Snippets (not state) can be saved as public, unlisted or secret on glot.io",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": "No GPU option is documented",
      "networking": "Network disabled on the hosted site; DOCKER_CONTAINER_NETWORK_DISABLED is a required setting when self-hosting",
      "pricing": "No pricing published; glot.io is free to use and the components are open source",
      "free_tier": "glot.io is free; free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "A free playground and a minimal MIT runner for stateless, network-less snippet execution",
      "overview": [
        "glot.io is an open-source code playground where you run and share snippets in 44 languages. The system is split into repositories under the glotcode organisation: `glot-app` (the website), `glot-languages` (language definitions and Docker images) and `docker-run`, the HTTP API that does the execution. docker-run talks to the Docker daemon over its Unix socket, writes the payload to a fresh container's stdin and reads a JSON result with stdout, stderr and error; every request gets a single-use container that is deleted afterwards, with cleanup handled by background workers and a prewarm pool that still never reuses a container.",
        "Operators set the limits through environment variables: network disabled, memory in bytes, nproc and nofile ulimits, dropped capabilities, a read-only root filesystem, a maximum execution time and a maximum output size. Runs are stateless. The hosted run API was documented in the author's `glot-run` repositories, which are now deprecated or archived (glot-run2 archived 26 August 2026, with a note that it would be incorporated into glot-www); API documentation, rate limits and pricing are not published on glot.io. docker-run is MIT and was updated in August 2026; the `glot` overview repository carries no licence file."
      ],
      "notable": [
        "44 languages listed on glot.io; code runs in a temporary container with network disabled",
        "docker-run (MIT, about 120 stars) last pushed August 2026; glot-run2 archived 26 August 2026",
        "The glot overview repository has about 1.7k stars but no licence file"
      ],
      "pros": [
        "You want a free, open-source playground to share runnable snippets",
        "You need a minimal MIT runner (docker-run) that guarantees a fresh container per request",
        "You are fine with network-less, stateless runs and want to set your own time, memory and process limits"
      ],
      "cons": [
        "Hosted API documentation, rate limits, an SLA and pricing are not published on glot.io; the glot-run API repositories are deprecated or archived",
        "Stateless and network-less by design",
        "Isolation is a Docker container (the gVisor runtime is optional when self-hosting); a CPU limit is not among the documented settings",
        "The glot overview repository was last pushed in August 2024"
      ],
      "verdict": "A free code playground and an MIT-licensed single-use container runner (docker-run) for stateless snippet execution; documentation for a hosted third-party API is not published on glot.io.",
      "peers": [
        "codapi",
        "piston",
        "judge0",
        "onecompiler",
        "jdoodle"
      ],
      "sources": [
        "https://glot.io",
        "https://github.com/glotcode/glot",
        "https://github.com/glotcode/docker-run",
        "https://github.com/glotcode/docker-run/blob/main/api_docs/run.md",
        "https://github.com/prasmussen/glot-run2",
        "https://github.com/prasmussen/glot-run",
        "https://github.com/glotcode/docker-run/blob/main/LICENSE"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "datalayer",
      "name": "Datalayer Runtimes",
      "company": "Datalayer",
      "hq": null,
      "kind": "hosted",
      "tier": 2,
      "alternatives": true,
      "website": "https://datalayer.ai",
      "docs": "https://core.datalayer.tech/python/",
      "pricing_url": "https://datalayer.ai/pricing",
      "github": "https://github.com/datalayer/core",
      "tagline": "Remote Jupyter kernels as a service for AI agents: stateful CPU or GPU runtimes created on demand from Python or TypeScript, with snapshots of installed packages, files and variables, and a hosted Jupyter MCP Server.",
      "isolation": "Not documented; the SDK describes each runtime as \"a clean, isolated Python environment\" on cloud machines and the agent-runtimes README refers to a pod's egress boundary, but no hypervisor or container technology is named",
      "isolation_short": "Not documented",
      "isolation_class": null,
      "open_source": null,
      "license": null,
      "self_host": null,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "CLI",
        "MCP server",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Runtimes are created with a timeout in minutes (SDK example: 30); the default and maximum are not stated on the SDK docs",
      "max_session_short": "Minutes (set per runtime)",
      "persistence": "Stateful kernel: variables persist across execute() calls within a runtime session and are lost on termination; snapshots capture installed packages, files and variables and can seed a new runtime",
      "snapshots": true,
      "pause_resume": null,
      "gpu": true,
      "gpu_detail": "Yes: an \"AI Environment\" GPU environment alongside the Python CPU environment; the pricing page lists A100 (80 GB) and T4 (16 GB) environments, shown as \"Unavailable\" when checked",
      "networking": null,
      "pricing": "Plans on datalayer.ai/pricing: Free (50 wallet credits and 500 runs per month), Team $49/month (500 wallet credits and 5,000 runs per month, plus $15 per active seat for organisations) and Enterprise (custom). Sandboxes consume wallet credits while running; the 1 CPU environment (1 vCPU, 1 GiB) is listed at 0.3 USD per hour",
      "free_tier": "Free plan: 50 wallet credits and 500 runs per month",
      "compliance": "Trust page states SOC 2 Type II, ISO/IEC 27001:2022 and ISO/IEC 42001:2023",
      "regions": "Available environments \"may vary based on your account tier and region\"; regions not listed",
      "best_for": "Agents that need a stateful, snapshot-able Jupyter kernel (CPU or GPU) rather than a one-shot run",
      "overview": [
        "Datalayer sells Jupyter-based compute for AI agents. A \"Runtime\" is a remote kernel you create on demand from the `datalayer_core` Python SDK (`client.create_runtime(environment=..., timeout=...)`), the TypeScript client or the CLI, then call `execute()` on; variables persist between calls within the session, which makes it closer to a code interpreter than to an online judge. Environments are preset: a Python CPU environment and a GPU \"AI Environment\", each with its own CPU, memory and GPU limits and an hourly burning rate. Secrets are injected as environment variables. Datalayer also hosts the open-source Jupyter MCP Server (BSD-3-Clause, about 1.3k stars) at mcp.datalayer.run so MCP clients such as Claude Code, Codex and Cursor can drive notebooks and sandboxes, and ships the BSD-3-Clause `agent-runtimes` server for hosting agents on Datalayer runtimes or locally.",
        "Runtimes are time-boxed (timeout in minutes; default not stated on the SDK page) and the environments are described as stateless, so persistent work goes into named snapshots that capture installed packages, files and in-memory variables and can be restored into a new runtime. The isolation mechanism is not documented on the pages checked and regions are not listed; pricing (Free, Team at $49/month, Enterprise custom) and compliance attestations (SOC 2 Type II, ISO/IEC 27001:2022, ISO/IEC 42001:2023) are published on the JavaScript-rendered datalayer.ai site; the SDKs (BSD-3-Clause) were still being pushed in November 2026."
      ],
      "notable": [
        "datalayer_core SDK 1.2.29 on PyPI (BSD-3-Clause); core and agent-runtimes repos pushed on 14 November 2026",
        "Hosted Jupyter MCP Server at mcp.datalayer.run with OAuth 2.1 sign-in and a Claude Code plugin",
        "Snapshots capture packages, files and variables but not active network connections"
      ],
      "pros": [
        "You want a stateful Jupyter kernel an agent can call repeatedly, with GPU environments",
        "You need to snapshot an environment (packages, files, variables) and restore it later",
        "You are driving notebooks from an MCP client and want a hosted MCP server"
      ],
      "cons": [
        "Isolation technology and regions are not documented on the pages checked",
        "Runtimes are time-boxed in minutes and environments are stateless without snapshots",
        "On-premise deployment is listed only on the Enterprise plan"
      ],
      "verdict": "Kernels-as-a-service with stateful runtimes and snapshots; the isolation technology behind its runtimes is not documented on the pages checked.",
      "peers": [
        "jupyter-kernel-gateway",
        "e2b",
        "modal",
        "lightning-ai-sandboxes",
        "openai-code-interpreter"
      ],
      "sources": [
        "https://datalayer.ai",
        "https://core.datalayer.tech/python/",
        "https://core.datalayer.tech/python/Runtimes",
        "https://core.datalayer.tech/python/Environments",
        "https://github.com/datalayer/core",
        "https://github.com/datalayer/agent-runtimes",
        "https://github.com/datalayer/jupyter-mcp-server",
        "https://pypi.org/pypi/datalayer_core/json",
        "https://datalayer.ai/pricing",
        "https://datalayer.ai/trust",
        "https://datalayer.ai/features/notebooks"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "jupyter-kernel-gateway",
      "name": "Jupyter Kernel Gateway",
      "company": "Project Jupyter",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://jupyter-kernel-gateway.readthedocs.io/en/latest/",
      "docs": "https://jupyter-kernel-gateway.readthedocs.io/en/latest/",
      "pricing_url": null,
      "github": "https://github.com/jupyter-server/kernel_gateway",
      "tagline": "BSD-licensed headless web server that exposes Jupyter kernels over REST and WebSockets; on its own it runs kernels as local processes, while the sibling Enterprise Gateway launches each kernel in its own Kubernetes pod and namespace.",
      "isolation": "Not documented for Kernel Gateway itself, which launches kernels in its own local process and filesystem space; the docs say it \"can be containerized and scaled out\". Enterprise Gateway launches each kernel as a Kubernetes pod (or on YARN, Docker Swarm or IBM Spectrum Conductor), by default in a per-kernel namespace with the baseline Pod Security Standard and root prohibited",
      "isolation_short": "Process (KG) / pod (EG)",
      "isolation_class": "container",
      "open_source": true,
      "license": "BSD-3-Clause",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "REST API",
        "Python"
      ],
      "cold_start": "Enterprise Gateway kernel launch timeout defaults to 60 seconds on Kubernetes",
      "cold_start_short": null,
      "max_session": "No limit by default; Enterprise Gateway idle culling is off by default (cull_idle_timeout 0, with 43,200 s recommended) and the Kubernetes Helm chart defaults to 3,600 s",
      "max_session_short": "Unbounded (culling optional)",
      "persistence": "Stateful kernel: variables and files persist for the life of the kernel; Kernel Gateway can seed every kernel from a notebook (seed_uri) and prespawn kernels; no snapshots",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Enterprise Gateway accepts KERNEL_GPUS / KERNEL_GPUS_LIMIT requests on Kubernetes; GPU handling is not documented for Kernel Gateway itself",
      "networking": "Not restricted by the gateways; whatever the host or pod allows",
      "pricing": "Free (BSD-3-Clause)",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Teams already on Jupyter and Kubernetes who want to run their own stateful kernels-as-a-service",
      "overview": [
        "Jupyter Kernel Gateway is a web server from Project Jupyter that gives applications headless access to Jupyter kernels: start and stop kernels through a REST API compatible with the Notebook server's kernel endpoints, then send code over a WebSocket carrying the Jupyter messaging protocol; a second `notebook-http` mode turns annotated notebook cells into HTTP endpoints. Configuration covers an auth token, `max_kernels`, a `seed_uri` notebook run on every kernel launch and `prespawn_count`. It documents no sandboxing of its own (kernels run in the gateway's process and filesystem space); the docs describe running it from a docker-stacks image and scaling it out with Kubernetes. The latest release is 3.0.1 (12 March 2024).",
        "Jupyter Enterprise Gateway extends the same idea into a multi-tenant \"kernel as a service\" that launches each kernel on a resource manager: on Kubernetes every kernel becomes its own pod (image chosen per kernelspec or via KERNEL_IMAGE from an allow-list), by default in a freshly created namespace with the baseline Pod Security Standard, non-root UID 1000, CPU/memory/GPU requests and limits from the client, and optional idle culling; YARN, Docker Swarm and IBM Spectrum Conductor are also supported. Version 3.4.0 shipped on 13 November 2026. Both are BSD-3-Clause. It gives you stateful kernels; on Kubernetes, isolation follows the namespace and Pod Security settings you configure."
      ],
      "notable": [
        "Kernel Gateway 3.0.1 released 12 March 2024; Enterprise Gateway 3.4.0 released 13 November 2026",
        "Enterprise Gateway creates one Kubernetes pod per kernel in a per-kernel namespace by default, with root prohibited and the baseline Pod Security Standard applied (since 3.4.0 configurable via EG_KERNEL_NAMESPACE_POD_SECURITY)",
        "Enterprise Gateway idle culling defaults to off (0 s) in the config docs, 3,600 s in the Helm chart"
      ],
      "pros": [
        "You want stateful Jupyter kernels (Python, R, Scala) under your own control, free of licence cost",
        "You run Kubernetes and want one pod per kernel with per-kernel namespaces and resource limits",
        "You are building on the standard Jupyter protocol so any kernel language and client works"
      ],
      "cons": [
        "Kernel Gateway launches kernels in its own process and filesystem space; isolation between kernels is not documented",
        "Enterprise Gateway warns that KERNEL_USERNAME comes from the client, so per-user namespaces are not a full access control",
        "Snapshots, pause/resume and per-call time limits are not documented; sessions are bounded by optional idle culling",
        "Kernel Gateway has had no release since March 2024"
      ],
      "verdict": "An open-source building block for kernels-as-a-service: stateful kernels over REST and WebSockets, with isolation provided by the container or Kubernetes pod security settings you deploy it with.",
      "peers": [
        "datalayer",
        "gke-agent-sandbox",
        "microsandbox",
        "openhands-runtime",
        "kata-containers"
      ],
      "sources": [
        "https://github.com/jupyter-server/kernel_gateway",
        "https://jupyter-kernel-gateway.readthedocs.io/en/latest/",
        "https://jupyter-kernel-gateway.readthedocs.io/en/latest/config-options.html",
        "https://jupyter-kernel-gateway.readthedocs.io/en/latest/websocket-mode.html",
        "https://pypi.org/project/jupyter-kernel-gateway/",
        "https://github.com/jupyter-server/enterprise_gateway",
        "https://jupyter-enterprise-gateway.readthedocs.io/en/latest/",
        "https://jupyter-enterprise-gateway.readthedocs.io/en/latest/operators/deploy-kubernetes.html",
        "https://jupyter-enterprise-gateway.readthedocs.io/en/latest/operators/config-culling.html",
        "https://pypi.org/pypi/jupyter_enterprise_gateway/json"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "firecracker",
      "name": "Firecracker",
      "company": "Amazon Web Services",
      "hq": {
        "city": "Seattle",
        "region": "Washington",
        "country": "US",
        "source": "https://aws.amazon.com/privacy/"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://firecracker-microvm.github.io",
      "docs": "https://github.com/firecracker-microvm/firecracker/tree/main/docs",
      "pricing_url": null,
      "github": "https://github.com/firecracker-microvm/firecracker",
      "tagline": "The KVM-based microVM monitor from AWS that underpins Lambda and Fargate, and the sandboxes of E2B, Vercel, Fly.io, Koyeb and Northflank.",
      "isolation": "KVM-based microVM with its own guest kernel; the VMM exposes a minimal device set (virtio-net, virtio-block, virtio-vsock, serial console, keyboard controller) and a companion jailer adds cgroup, namespace and seccomp barriers",
      "isolation_short": "KVM microVM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "REST API",
        "CLI"
      ],
      "cold_start": "Specification target: 125 ms or less from the InstanceStart API call to guest /sbin/init, with the serial console disabled and a minimal kernel and root filesystem; VMM process start within 8 CPU ms",
      "cold_start_short": "<125 ms (spec)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Pause/resume through the /vm API; full snapshots serialise guest memory and device state so a microVM can be restored in another Firecracker process on an identical hardware and software setup; diff snapshots (dirty memory only) are in developer preview",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": "Not documented; the documented device set is virtio-net, virtio-block, virtio-vsock, entropy, pmem, balloon, serial console and keyboard controller",
      "networking": "virtio-net interfaces attached to host tap devices by the embedder, with built-in rate limiters for bandwidth and operations per second",
      "pricing": "Free and open source",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Building your own microVM sandbox fleet on Linux hosts with KVM",
      "overview": [
        "Firecracker is a virtual machine monitor written in Rust by developers at Amazon Web Services and released under Apache-2.0 (crosvm-derived sections under BSD-3-Clause). It runs in user space on a Linux host and uses KVM to create **microVMs**, each with its own guest kernel; the project describes itself as \"purpose-built for creating and managing secure, multi-tenant container and function-based services\". A Firecracker process exposes an OpenAPI-specified REST API over a socket through which you configure vCPUs, memory, block and network devices, then boot the guest. The companion **jailer** applies cgroup and namespace isolation and drops privileges as \"a second line of defense in case the virtualization barrier is ever compromised\". Hosts need 64-bit Intel, AMD or Arm CPUs with hardware virtualisation; guests are Linux kernels 4.14 and above.",
        "Firecracker is the primitive under much of the hosted sandbox market. AWS says it developed Firecracker for AWS Lambda and AWS Fargate, and the project site lists E2B, Fly.io, Koyeb, Northflank, Kata Containers, OpenNebula and Qovery among users and integrations. E2B states that \"Every session receives a hardware-isolated Firecracker microVM\", and Vercel documents that \"Each sandbox runs in a secure Firecracker microVM with its own filesystem and network\". Snapshot and restore are built in, which is what makes pause/resume and forking on those platforms possible. The open-source project is not itself a managed service: you supply hosts, kernels, root filesystems and an orchestrator, and the project site points to AWS Lambda MicroVMs as a managed alternative. Releases arrive roughly every two to three months; v1.18.0 was published on 9 November 2026."
      ],
      "notable": [
        "Developed at AWS for AWS Lambda and AWS Fargate; the project site says the same virtualisation sits behind \"15 trillion+ monthly Lambda function invocations\"",
        "Specification targets: VMM memory overhead of 5 MiB or less per microVM (1 vCPU, 128 MiB RAM) and guest boot of 125 ms or less; the site also claims \"Create up to 150 microVMs per second per host\"",
        "v1.18.0 released 9 November 2026; diff snapshots remain in developer preview"
      ],
      "pros": [
        "You want the same microVM boundary that E2B, Vercel Sandbox, Fly.io and AWS Lambda use, on your own hardware",
        "You need pause/resume and memory snapshots to fork or checkpoint sandboxes",
        "You run Linux hosts with KVM and can build your own orchestration layer",
        "You want a small, Rust-based VMM with a documented security model (jailer, seccomp, rate limiters)"
      ],
      "cons": [
        "Linux hosts only; the KVM backend needs hardware virtualisation on the host (bare metal or nested virtualisation), and the alternative Nitro Sandbox backend added in v1.18 runs only inside Amazon EC2 instances",
        "GPU passthrough is not documented",
        "A VMM only: you must provide guest kernels, root filesystems, networking and a control plane yourself",
        "Snapshots must be restored on an identical hardware and software configuration, and network connections are not guaranteed to survive a resume"
      ],
      "verdict": "An open-source microVM monitor from AWS that several sandbox platforms build on; it suits teams building a platform rather than consuming one.",
      "peers": [
        "kata-containers",
        "libkrun",
        "gvisor",
        "e2b",
        "fly-machines",
        "aws-lambda-microvms"
      ],
      "sources": [
        "https://firecracker-microvm.github.io",
        "https://github.com/firecracker-microvm/firecracker",
        "https://github.com/firecracker-microvm/firecracker/blob/main/SPECIFICATION.md",
        "https://github.com/firecracker-microvm/firecracker/blob/main/docs/snapshotting/snapshot-support.md",
        "https://github.com/firecracker-microvm/firecracker/releases",
        "https://e2b.dev",
        "https://vercel.com/docs/vercel-sandbox",
        "https://github.com/firecracker-microvm/firecracker/blob/main/docs/nitro-sandboxes/README.md",
        "https://github.com/firecracker-microvm/firecracker/blob/main/FAQ.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "kata-containers",
      "name": "Kata Containers",
      "company": "OpenInfra Foundation",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://katacontainers.io",
      "docs": "https://github.com/kata-containers/kata-containers/tree/main/docs",
      "pricing_url": null,
      "github": "https://github.com/kata-containers/kata-containers",
      "tagline": "An OCI runtime that runs each Kubernetes pod in a lightweight VM with its own kernel, on QEMU, Cloud Hypervisor, Firecracker, Dragonball or StratoVirt.",
      "isolation": "Lightweight VM per pod with a dedicated guest kernel, run by a pluggable hypervisor (QEMU, Cloud Hypervisor, Firecracker, Dragonball or StratoVirt) behind a containerd shim v2",
      "isolation_short": "VM per pod",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": null,
      "snapshots": null,
      "pause_resume": null,
      "gpu": true,
      "gpu_detail": "Yes - the hypervisors document says QEMU \"is the best supported hypervisor for NVIDIA-based GPUs and for confidential computing use-cases\"",
      "networking": null,
      "pricing": "Free and open source",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Kubernetes clusters that need VM isolation per pod without leaving the container toolchain",
      "overview": [
        "Kata Containers is an open-source container runtime stewarded by the OpenInfra Foundation and licensed Apache-2.0. It builds \"lightweight virtual machines that seamlessly plug into the containers ecosystem\": each workload \"runs in a dedicated kernel, providing isolation of network, I/O and memory\", with hardware virtualisation as \"a second layer of defense\". The main runtime is a containerd shim v2 implementation that speaks OCI and the Kubernetes CRI, so a cluster selects it with a RuntimeClass; packaging includes the kata-deploy Helm chart. Kata is a hypervisor orchestration layer rather than a VMM: its hypervisors document lists QEMU (all architectures, best for NVIDIA GPUs and confidential computing), Cloud Hypervisor, Firecracker, Ant Group's in-process Dragonball and StratoVirt (the last four on x86_64 and aarch64), and supported architectures include ppc64le and s390x.",
        "On this site, Northflank is the platform that documents building on Kata: its security page says \"Kata Containers are used as the default runtime class for workloads in the cluster\" and that this \"applies KVM and micro-VMs through Cloud Hypervisor to secure the workloads from the host kernel\", with gVisor as the alternative where nested virtualisation is unavailable. Kata itself names Baidu as a production user for function computing and container instances. Kata Containers 4.3.1 was released on 21 October 2026. Snapshots, pause/resume and session limits are not documented on the project pages checked."
      ],
      "notable": [
        "Stewarded by the OpenInfra Foundation; Apache-2.0",
        "Kata Containers 4.3.1 released 21 October 2026 (4.3.0 on 20 October, 4.2.0 on 15 September)",
        "Northflank documents Kata with Cloud Hypervisor as its default runtime class; Baidu runs Kata in production for Function Computing, Cloud Container Instances and Edge Computing"
      ],
      "pros": [
        "You run Kubernetes and want VM-grade isolation per pod selected with a RuntimeClass",
        "You need a choice of hypervisor, including QEMU for GPU and confidential-computing workloads",
        "You want an actively maintained, foundation-governed project with multi-architecture support"
      ],
      "cons": [
        "Needs hardware virtualisation on the node (Intel VT-x, AMD SVM, ARM Hyp, IBM Power or IBM Z SIE), which on cloud VMs means bare-metal or nested-virtualisation instances",
        "A container runtime rather than a sandbox API; snapshots and pause/resume are not documented at the project level",
        "Each pod involves a guest kernel, an agent, a containerd shim and a hypervisor"
      ],
      "verdict": "An open-source container runtime that runs each Kubernetes pod in a lightweight VM; it suits clusters whose sandboxes are already pods.",
      "peers": [
        "firecracker",
        "gvisor",
        "edera",
        "gke-agent-sandbox",
        "northflank"
      ],
      "sources": [
        "https://katacontainers.io",
        "https://github.com/kata-containers/kata-containers",
        "https://github.com/kata-containers/kata-containers/blob/main/docs/hypervisors.md",
        "https://github.com/kata-containers/kata-containers/releases",
        "https://northflank.com/security",
        "https://github.com/kata-containers/kata-containers/blob/main/README.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "gvisor",
      "name": "gVisor",
      "company": "Google",
      "hq": {
        "city": "Mountain View",
        "region": "California",
        "country": "US",
        "source": "https://policies.google.com/terms"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://gvisor.dev",
      "docs": "https://gvisor.dev/docs/",
      "pricing_url": null,
      "github": "https://github.com/google/gvisor",
      "tagline": "Google's user-space application kernel for containers, with the runsc OCI runtime, GPU passthrough via nvproxy and checkpoint/restore; used by Modal, Beam, Anthropic, OpenAI and Google Cloud.",
      "isolation": "User-space application kernel written in Go that intercepts the sandboxed application's system calls; default systrap platform (seccomp traps), optional KVM platform",
      "isolation_short": "User-space kernel",
      "isolation_class": "gvisor",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": "gvisor.dev claims containers \"start up in milliseconds and have minimal resource overhead\"; no figure given",
      "cold_start_short": "ms (claim)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Checkpoint a running container to a state file and restore it into a new container (runsc checkpoint / runsc restore), with compression, background restore and workload-triggered checkpoints; GPU checkpoint/restore is x86_64 only",
      "snapshots": true,
      "pause_resume": null,
      "gpu": true,
      "gpu_detail": "Yes - nvproxy runs \"most CUDA applications on preselected versions of NVIDIA's open source driver\"; T4, A100, A10G, L4 and H100 officially supported; gVisor \"doesn't introduce any additional hardware-level isolation\" beyond the NVIDIA kernel-mode driver's own",
      "networking": "Own user-space network stack written in Go (netstack); --network=host mode passes through host sockets",
      "pricing": "Free and open source",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Sandboxing containers on hosts where you cannot or do not want to run a VM per workload, including GPU workloads",
      "overview": [
        "gVisor is an open-source, Apache-2.0 \"application kernel for containers\" that originated at Google and \"leverages years of experience isolating production workloads at Google\". Rather than a seccomp filter or a VM, it implements a Linux-compatible kernel in Go that runs in user space and intercepts the application's system calls, so the host kernel sees only a narrow interface; the README summarises this as \"gVisor implements Linux by way of Linux\". It ships as an OCI runtime called **runsc** that plugs into Docker, containerd and Kubernetes. The default platform is systrap, which uses seccomp traps and \"replaced ptrace as the default gVisor platform in mid-2023\"; a KVM platform uses hardware virtualisation and \"runs best on bare-metal setups\". Hosts need Linux 5.6+ on x86_64 or arm64.",
        "The adopters file shows how widely sandbox platforms build on it: \"gVisor powers Google Cloud offerings GKE Sandbox, Cloud Run, App Engine\"; Modal says \"Compute jobs at Modal are containerized and virtualized using gVisor\"; Beam combines \"stateful snapshots, gVisor isolation, GPU acceleration\"; Anthropic uses it \"to securely contain code execution within claude.ai\"; OpenAI uses it for \"some higher-risk tasks\"; Northflank \"uses gVisor to run GPU workloads inside sandboxed environments\"; Cloudflare, DigitalOcean, Snowflake and Ant Group are also listed. Two features matter for agent sandboxes: **nvproxy** passes NVIDIA driver ioctls through so CUDA workloads run inside the sandbox, and **checkpoint/restore** saves a running container to a state file. Releases are dated tags, most recently release-20261109.0."
      ],
      "notable": [
        "Adopters page: powers GKE Sandbox, Cloud Run and App Engine; also lists Modal, Beam, Anthropic (claude.ai), OpenAI, Northflank, Cloudflare, DigitalOcean, Snowflake and Ant Group",
        "systrap became the default platform in mid-2023; ptrace is \"no longer supported and is expected to eventually be removed entirely\"",
        "Latest release tag release-20261109.0 (11 November 2026); Apache-2.0"
      ],
      "pros": [
        "You want a kernel boundary without hardware virtualisation, so it works inside ordinary cloud VMs",
        "You need GPU workloads sandboxed, as Modal, Beam and Northflank do with nvproxy",
        "You run Docker or Kubernetes and want to swap in runsc as a runtime class",
        "You want checkpoint/restore of running containers"
      ],
      "cons": [
        "Application system calls \"must traverse additional layers of software\", which the docs call a structural cost; the KVM platform \"runs best on bare-metal setups\"",
        "Only a limited set of NVIDIA ioctls and driver versions are supported, and gVisor \"doesn't introduce any additional hardware-level isolation\" for GPU driver bugs",
        "Not every Linux feature is implemented, so some applications need compatibility checks",
        "Linux x86_64 and arm64 hosts only"
      ],
      "verdict": "A user-space application kernel for containers whose adopters file lists Google Cloud and several sandbox platforms; it suits hosts where a VM per workload is impractical and workloads that need sandboxed GPUs.",
      "peers": [
        "kata-containers",
        "firecracker",
        "nsjail",
        "modal",
        "gke-agent-sandbox"
      ],
      "sources": [
        "https://gvisor.dev",
        "https://github.com/google/gvisor",
        "https://github.com/google/gvisor/blob/master/ADOPTERS.md",
        "https://gvisor.dev/docs/architecture_guide/platforms/",
        "https://gvisor.dev/docs/user_guide/gpu/",
        "https://gvisor.dev/docs/user_guide/checkpoint_restore/",
        "https://github.com/google/gvisor/releases",
        "https://modal.com/docs/guide/security",
        "https://gvisor.dev/docs/architecture_guide/performance/",
        "https://gvisor.dev/docs/user_guide/compatibility/",
        "https://gvisor.dev/docs/user_guide/install/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "libkrun",
      "name": "libkrun",
      "company": "libkrun project",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/libkrun/libkrun",
      "docs": "https://github.com/containers/libkrun/tree/main/docs",
      "pricing_url": null,
      "github": "https://github.com/libkrun/libkrun",
      "tagline": "A dynamic library that lets any program launch a microVM in-process, on KVM in Linux or Hypervisor.framework on macOS; used by crun/Podman, krunkit and microsandbox.",
      "isolation": "microVM embedded in the calling process via KVM on Linux or Hypervisor.framework (HVF) on macOS/ARM64, with a minimal set of virtio devices; the README calls this a \"partially isolated environment\"",
      "isolation_short": "In-process microVM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Rust",
        "C"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": null,
      "snapshots": null,
      "pause_resume": null,
      "gpu": true,
      "gpu_detail": "Yes - virtio-gpu with venus and native-context; krunkit uses it for \"GPU-enabled (via venus) lightweight VMs on macOS\" and muvm for GPU-accelerated microVMs",
      "networking": "Transparent Socket Impersonation (TSI) or virtio-net connected to a user-space proxy such as passt or gvproxy",
      "pricing": "Free and open source",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Embedding a microVM in your own tool on Linux or Apple Silicon without shipping a separate VMM",
      "overview": [
        "libkrun is an Apache-2.0 dynamic library hosted in the libkrun GitHub organisation, previously under containers (its security contact is a redhat.com mailing list) that gives a program \"the ability to run processes in a partially isolated environment\" using \"KVM Virtualization on Linux, HVF on macOS/ARM64\", with Windows Hypervisor Platform support listed as upcoming. It bundles a VMM with virtio console, block, fs, gpu, net, vsock, balloon, rng and rtc devices and exposes a native Rust API plus an autogenerated C API, so the microVM is created inside the calling process rather than by launching a separate binary. Variants add AMD SEV (libkrun-sev) and Intel TDX (libkrun-tdx) memory encryption. The README warns that \"The VM boundary does not restrict host resources exposed through devices\".",
        "libkrun is the microVM layer in several tools: crun uses it for \"virtualization-based isolation to container and confidential workloads\" (Podman's --runtime=krun), krunkit runs GPU-enabled VMs on macOS, and muvm launches GPU-accelerated microVMs for games. microsandbox, which is on this site, credits libkrun in its README acknowledgements as a project that \"made microsandbox possible\" alongside smoltcp. Snapshot, session and orchestration features are not documented for the library. The stable branch reached libkrun-1.19.8 on 6 November 2026; the main branch is libkrun 2.0, which \"will not be backwards compatible with the 1.x API/ABI\" and whose API \"may change further before the first stable release\"."
      ],
      "notable": [
        "Runs on Linux (KVM) and macOS/ARM64 (Hypervisor.framework), the only microVM primitive in this batch that documents Apple Silicon",
        "libkrun-1.19.8 released 6 November 2026; main branch is libkrun 2.0 with a changing API",
        "Named users: crun (Podman --runtime=krun), krunkit and muvm; microsandbox credits it in its acknowledgements"
      ],
      "pros": [
        "You want microVM isolation on a developer Mac as well as Linux servers",
        "You are writing your own sandbox tool and want to embed the VMM as a library with Rust or C bindings",
        "You need GPU access inside the VM via virtio-gpu",
        "You want confidential-computing variants (SEV, TDX) from the same codebase"
      ],
      "cons": [
        "Described by its own README as \"partially isolated\"; it cautions that \"The VM boundary does not restrict host resources exposed through devices\"",
        "No snapshot/restore or pause/resume documented",
        "Rust and C APIs only; a higher-level SDK or control plane is not documented",
        "Windows support is upcoming, not shipped"
      ],
      "verdict": "A library for embedding microVM-based isolation in your own tool on Linux (KVM) or macOS/ARM64 (Hypervisor.framework).",
      "peers": [
        "firecracker",
        "microsandbox",
        "kata-containers",
        "apple-container",
        "smolvm"
      ],
      "sources": [
        "https://github.com/libkrun/libkrun",
        "https://github.com/libkrun/libkrun/blob/main/README.md",
        "https://github.com/libkrun/libkrun/releases",
        "https://github.com/zerocore-ai/microsandbox",
        "https://github.com/libkrun/libkrun/blob/main/docs/security-model.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "bubblewrap",
      "name": "bubblewrap",
      "company": "containers project",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/containers/bubblewrap",
      "docs": "https://github.com/containers/bubblewrap/blob/main/README.md",
      "pricing_url": null,
      "github": "https://github.com/containers/bubblewrap",
      "tagline": "The unprivileged Linux namespace sandboxer behind Flatpak and Anthropic's sandbox runtime; you assemble the policy from command-line flags.",
      "isolation": "Unprivileged Linux user, mount, PID, network, IPC and UTS namespaces with a tmpfs root, PR_SET_NO_NEW_PRIVS and optional seccomp filters; the caller defines the policy",
      "isolation_short": "Linux namespaces",
      "isolation_class": "os",
      "open_source": true,
      "license": "LGPL-2.1",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "None; the sandbox starts from an empty tmpfs root that is cleaned up when the last process exits, with host paths bind-mounted as you choose",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Host network by default; --unshare-net gives a network namespace with loopback only",
      "pricing": "Free and open source",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Lightweight process sandboxing on Linux where you control the policy and do not need a kernel boundary",
      "overview": [
        "bubblewrap (bwrap) is an LGPL-2.1 \"low-level sandboxing tool for running programs without root privileges\" from the containers GitHub organisation, with roots in the xdg-app helper and linux-user-chroot. It starts each run in a new mount namespace with an empty tmpfs root and bind-mounts only the host paths you pass; user namespaces hide every uid and gid except the current one; PID, IPC, network and UTS namespaces can be unshared; seccomp filters can be supplied; and PR_SET_NO_NEW_PRIVS disables setuid binaries. The README is explicit that \"bubblewrap is not a complete, ready-made sandbox with a specific security policy\" and that \"Everything mounted into the sandbox can potentially be used to escalate privileges\", so callers are expected to be careful with sockets such as D-Bus. The old setuid mode has been removed; unprivileged user namespaces are required.",
        "bubblewrap is the sandbox under Flatpak, rpm-ostree's unprivileged mode and bwrap-oci, and it is the Linux half of Anthropic's sandbox runtime (anthropic-srt on this site), whose README says it \"Uses bubblewrap for containerization with network namespace isolation\" while using sandbox-exec on macOS. It is a single binary with no daemon, no API and no persistence; everything is a command-line invocation, which makes it easy to wrap but leaves orchestration, limits and policy to you. bubblewrap 0.13.1 was released on 14 October 2026."
      ],
      "notable": [
        "Used by Flatpak and by Anthropic's sandbox runtime on Linux",
        "bubblewrap 0.13.1 released 14 October 2026; LGPL-2.1",
        "Setuid mode removed; requires unprivileged user namespaces on the host"
      ],
      "pros": [
        "You want a tiny, well-understood tool to confine a coding agent's shell on a Linux machine",
        "You need unprivileged sandboxing without a daemon, container runtime or hypervisor",
        "You are building a local agent sandbox and want the same primitive Anthropic uses on Linux"
      ],
      "cons": [
        "Isolation uses host-kernel namespaces and optional seccomp filters; there is no separate guest kernel",
        "No built-in policy: the README warns that anything you mount in can be an escalation path",
        "Linux only, and the host must allow unprivileged user namespaces",
        "Snapshots, resource limits and lifecycle management are not documented"
      ],
      "verdict": "A low-level, unprivileged namespace sandboxing tool for Linux, suited to callers who define their own policy; it uses host-kernel namespaces rather than a separate guest kernel.",
      "peers": [
        "nsjail",
        "firejail",
        "anthropic-srt",
        "gvisor",
        "nono"
      ],
      "sources": [
        "https://github.com/containers/bubblewrap",
        "https://github.com/containers/bubblewrap/releases",
        "https://github.com/anthropic-experimental/sandbox-runtime",
        "https://github.com/containers/bubblewrap/blob/main/README.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "nsjail",
      "name": "nsjail",
      "company": "Google (not an official Google product)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://nsjail.dev",
      "docs": "https://github.com/google/nsjail/blob/master/README.md",
      "pricing_url": null,
      "github": "https://github.com/google/nsjail",
      "tagline": "A Linux process isolation tool combining namespaces, cgroups, rlimits and Kafel seccomp-bpf policies, with listen, once and rerun modes for untrusted programs.",
      "isolation": "Linux UTS, mount, PID, IPC, network, user, cgroup and time namespaces, cgroup v1/v2 limits, rlimits, chroot/pivot_root and seccomp-bpf syscall filters written in Kafel",
      "isolation_short": "Namespaces + seccomp",
      "isolation_class": "os",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Configurable CPU-time rlimit and wall-clock time limit per jail",
      "max_session_short": "Configurable",
      "persistence": "None built in; filesystem is whatever you chroot or bind-mount",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Network namespace with isolated or cloned Ethernet interfaces, MACVLAN (requires root) or userland networking via pasta",
      "pricing": "Free and open source",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Running untrusted programs or network services under tight syscall, memory and CPU limits on Linux",
      "overview": [
        "nsjail is an Apache-2.0 \"Linux process isolation tool using namespaces, resource limits, and seccomp-bpf syscall filters\" published under Google's GitHub organisation with the note that \"This is not an official Google product\". It isolates UTS, mount, PID, IPC, network, user, cgroup and time namespaces, constrains the filesystem with chroot or pivot_root and read-only bind mounts, caps CPU time, memory, file descriptors and process count through rlimits and cgroups (v1 and v2), and filters system calls with policies written in the Kafel language. Configuration is by command-line flags or Protocol Buffers config files. Operating modes include a listen mode (inetd-style TCP service), once mode, direct execve mode and a rerun mode for fuzzing.",
        "The documented use cases are hosting CTF challenges, fuzzing, sandboxing desktop applications such as Firefox and running programs in minimal environments with only a few read-only paths exposed; the same shape (bounded CPU, memory and syscalls per execution) is what code-execution APIs need. Some features need root (MACVLAN, --disable_clone_newuser) and others depend on unprivileged user namespaces; time namespaces need kernel 5.3 or later. nsjail 3.6 was released on 18 March 2026; its release notes say it \"Fixed seccomp policy assembly so repeated seccomp_string entries are combined and compiled together\", and the 3.5 release notes (5 March 2026) are marked \"BROKEN - seccomp_string(s) from config files are not additive\". None of the platforms on this site document building on nsjail."
      ],
      "notable": [
        "Hosted under Google's GitHub organisation but \"not an official Google product\"; Apache-2.0",
        "nsjail 3.6 released 18 March 2026; the 3.5 release notes (5 March 2026) are marked \"BROKEN - seccomp_string(s) from config files are not additive\"",
        "Seccomp policies are written in Kafel; time namespaces need kernel 5.3+, core scheduling 5.14+"
      ],
      "pros": [
        "You need per-execution CPU, memory, process and syscall limits for an untrusted code runner",
        "You want a single binary with a protobuf config rather than a container runtime",
        "You run network-facing untrusted services and want inetd-style listen mode"
      ],
      "cons": [
        "Isolation uses host-kernel namespaces, cgroups and seccomp-bpf filters; there is no separate guest kernel",
        "Linux only, and several features require root or enabled user namespaces",
        "Snapshots, persistence and an orchestration API are not documented; it is a command-line tool",
        "Release 3.5 (5 March 2026) came 29 months after 3.4 (October 2023) and its notes are marked \"BROKEN\" for a seccomp_string issue; 3.6 followed on 18 March 2026"
      ],
      "verdict": "An OS-level process jail for CTF challenges, fuzzing and other untrusted programs on Linux; it uses host-kernel namespaces and seccomp-bpf rather than a separate guest kernel.",
      "peers": [
        "bubblewrap",
        "firejail",
        "gvisor",
        "judge0",
        "piston"
      ],
      "sources": [
        "https://github.com/google/nsjail",
        "https://github.com/google/nsjail/blob/master/README.md",
        "https://nsjail.dev",
        "https://github.com/google/nsjail/releases",
        "https://github.com/google/nsjail/releases/tag/3.5",
        "https://github.com/google/nsjail/releases/tag/3.6"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "firejail",
      "name": "Firejail",
      "company": "netblue30 (community project)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://firejail.wordpress.com",
      "docs": "https://github.com/netblue30/firejail/wiki",
      "pricing_url": null,
      "github": "https://github.com/netblue30/firejail",
      "tagline": "A SUID Linux sandbox that confines desktop and server applications with namespaces, seccomp-bpf and capabilities, shipping around 1,350 application profiles.",
      "isolation": "SUID sandbox using Linux namespaces, seccomp-bpf and Linux capabilities, with per-application profiles and optional AppArmor and experimental Landlock integration",
      "isolation_short": "Namespaces + seccomp",
      "isolation_class": "os",
      "open_source": true,
      "license": "GPL-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "None built in; profiles control which home and system paths are visible or private",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Each sandboxed process tree gets its own view of the network stack via namespaces; profiles can restrict or disable networking",
      "pricing": "Free and open source",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Confining a browser or desktop application on a Linux workstation with ready-made profiles",
      "overview": [
        "Firejail is a GPL-2.0 \"lightweight security tool intended to protect a Linux system by setting up a restricted environment\" for running untrusted applications. It is \"an SUID sandbox program\" that uses \"Linux namespaces, seccomp-bpf and Linux capabilities\" so that each sandboxed process tree gets its own view of the network stack, process table and mount table, and it can run alongside SELinux or AppArmor; Landlock support is marked experimental. You start a sandbox by prefixing a command with firejail, and firecfg integrates \"over 900 default applications\" with the desktop. The project is written in C with \"virtually no dependencies\" and runs \"on any Linux computer with a 3.x kernel version or newer\".",
        "Unlike bubblewrap and nsjail, Firejail ships policy: its profile statistics report about 1,350 profiles, and the project site says sensitive data such as passwords and keys is blocked for more than 1,000 desktop applications. That makes it the quickest way to confine a browser an agent drives on a Linux desktop, but it is documented as a command-line tool for confining individual applications: an API, snapshotting and orchestration are not documented, and it runs as a SUID program whereas bubblewrap relies on unprivileged user namespaces. Firejail 0.9.82.2 was released on 3 November 2026. No platform on this site documents building on Firejail."
      ],
      "notable": [
        "About 1,350 application profiles; desktop integration for over 900 applications via firecfg",
        "Firejail 0.9.82.2 released 3 November 2026 (0.9.82 on 2 November); GPL-2.0",
        "Runs on any Linux with a 3.x or newer kernel; Landlock support is experimental"
      ],
      "pros": [
        "You want ready-made profiles for browsers and desktop applications rather than writing policy",
        "You run a coding agent or browser on a Linux workstation and want a one-word prefix to confine it",
        "You need something with no daemon and almost no dependencies"
      ],
      "cons": [
        "Runs as a SUID program (the README calls it \"an SUID sandbox program\") rather than through unprivileged user namespaces alone",
        "Isolation uses host-kernel namespaces, seccomp-bpf and capabilities; there is no separate guest kernel",
        "An API, snapshots and orchestration are not documented; sandboxes are started per command",
        "Linux only"
      ],
      "verdict": "A profile-driven SUID sandbox for confining individual applications on a Linux machine, with ready-made profiles for desktop software; an API and orchestration are not documented.",
      "peers": [
        "bubblewrap",
        "nsjail",
        "anthropic-srt",
        "gvisor",
        "nono"
      ],
      "sources": [
        "https://github.com/netblue30/firejail",
        "https://firejail.wordpress.com",
        "https://github.com/netblue30/firejail/releases",
        "https://github.com/netblue30/firejail/blob/master/README.md",
        "https://github.com/netblue30/firejail/blob/master/SECURITY.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "edera",
      "name": "Edera",
      "company": "Edera",
      "hq": {
        "city": "Seattle",
        "region": "Washington",
        "country": "US",
        "source": "https://edera.dev/privacy-policy"
      },
      "kind": "software",
      "tier": 2,
      "alternatives": true,
      "website": "https://edera.dev",
      "docs": "https://docs.edera.dev",
      "pricing_url": null,
      "github": "https://github.com/edera-dev",
      "tagline": "A commercial Kubernetes runtime that boots each pod into its own paravirtualised \"zone\" with a dedicated kernel on a Xen-derived hypervisor, without needing nested virtualisation.",
      "isolation": "Xen-derived microkernel hypervisor with control components rewritten in Rust; each pod runs in a zone (a lightweight VM with its own Linux kernel) as a paravirtualised guest, with PVH and a KVM backend as options",
      "isolation_short": "Xen-based zone VM",
      "isolation_class": "microvm",
      "open_source": false,
      "license": null,
      "self_host": true,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": null,
      "snapshots": null,
      "pause_resume": null,
      "gpu": true,
      "gpu_detail": "Yes - NVIDIA and AMD GPU passthrough, NVIDIA GPU Operator with zones and a DRA driver; GPU drivers can run in separate driver zones",
      "networking": null,
      "pricing": "Not published; the site offers \"Book a Meeting\" and early access forms",
      "free_tier": null,
      "compliance": null,
      "regions": null,
      "best_for": "Kubernetes clusters on ordinary cloud instances that need a kernel per pod without nested virtualisation",
      "overview": [
        "Edera is a venture-backed company whose \"Edera Hardened Runtime\" runs \"each Kubernetes pod with its own kernel for true hardware isolation at performance within 5% of native\". The architecture is a \"microkernel hypervisor (Xen)\" whose control components are \"rewritten in Rust for safety and maintainability\"; the hypervisor is limited to essential hardware interaction while networking and device drivers live in user-space processes in a hardened root zone or in separate **driver zones**, so \"Compromise of a driver does not compromise the host\". Each workload runs in a zone, \"a virtual machine guest environment where pods run with a dedicated zone kernel\" pulled as an OCI image. Zones run as paravirtualised guests, which the docs say avoids the need for nested virtualisation and lets Edera run on most cloud instances; PVH and an \"Edera KVM\" backend (early access) are alternatives. Pods opt in with runtimeClassName: edera.",
        "Install guides cover AWS EKS, Google Compute Engine, Azure Linux, Linode and Bottlerocket, plus a standalone mode without Kubernetes. The product is commercial and licensed under Edera's Software License Agreement, which describes it as proprietary software; the edera-dev GitHub organisation publishes components under open licences, notably **krata**, a GPL-2.0 Xen control plane described as \"the base layer upon which Edera Protect zones are built\", plus styrolite (Apache-2.0 programmatic sandboxing) and am-i-isolated. The product page claims \"Edera pods ran ~20% faster and up to 40% more efficient than Kata\" in one customer environment. Pricing and SLA terms are not published; the Trust Center lists a SOC 2 Type 1 (2026) report available on request. No sandbox platform on this site documents building on Edera."
      ],
      "notable": [
        "Backers listed on the site: 645 Ventures, MANTIS, Eniac, IQT, FPV and M12 (Microsoft's venture fund)",
        "krata, the open-source Xen control plane under Edera zones, is GPL-2.0; styrolite and other components are Apache-2.0",
        "Claims performance \"within 5% of native\" and ~20% faster than Kata in a customer environment (vendor figures)"
      ],
      "pros": [
        "You run Kubernetes on cloud instances without nested virtualisation and still want a kernel per pod",
        "You need GPU workloads isolated with drivers kept out of the host kernel",
        "You want VM isolation adopted through a standard RuntimeClass with no application changes"
      ],
      "cons": [
        "The runtime itself is proprietary with unpublished pricing; only components such as krata are open source",
        "Operates at the Kubernetes runtime layer with \"no SDK\"; snapshots and a per-execution API are not documented",
        "Built on a Xen-derived hypervisor rather than KVM; an \"Edera KVM\" backend is in early access",
        "SLA terms are not published; a SOC 2 Type 1 (2026) report is listed in the Trust Center on request"
      ],
      "verdict": "A commercial alternative to Kata for pod-level VM isolation on everyday cloud instances; evaluate it if nested virtualisation is your blocker.",
      "peers": [
        "kata-containers",
        "gvisor",
        "gke-agent-sandbox",
        "northflank",
        "firecracker"
      ],
      "sources": [
        "https://edera.dev",
        "https://edera.dev/product",
        "https://docs.edera.dev",
        "https://docs.edera.dev/technical-overview/architecture/overview/",
        "https://docs.edera.dev/technical-overview/concepts/zone/",
        "https://github.com/edera-dev",
        "https://github.com/edera-dev/krata",
        "https://edera.dev/software-license-agreement",
        "https://trust.edera.dev/",
        "https://edera.dev/use-case/ai-agent-sandboxing"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "wasmtime",
      "name": "Wasmtime",
      "company": "Bytecode Alliance",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://wasmtime.dev",
      "docs": "https://docs.wasmtime.dev",
      "pricing_url": null,
      "github": "https://github.com/bytecodealliance/wasmtime",
      "tagline": "The Bytecode Alliance's standalone WebAssembly and WASI runtime, embeddable from Rust, C, Python, Go, .NET and Ruby, and the engine under Spin, Extism, Akamai Functions and Fastly.",
      "isolation": "WebAssembly linear-memory sandbox with bounds-checked memory, 2 GB guard regions, capability-based WASI file access, memory zeroing between instances and fuel or epoch-based execution interruption",
      "isolation_short": "Wasm sandbox",
      "isolation_class": "wasm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "CLI",
        "Rust",
        "C",
        "C++",
        "Python",
        ".NET",
        "Go",
        "Ruby"
      ],
      "cold_start": "wasmtime.dev says it \"is optimized for efficient instantiation\" with scalable concurrent instances; no figure given",
      "cold_start_short": null,
      "max_session": "Embedder-controlled: fuel budgets (deterministic) or epoch deadlines (about 10% overhead) trap or yield the guest",
      "max_session_short": "Embedder-set",
      "persistence": "Not documented on the pages checked; the embedding host owns instance state and can pause a guest via async yield",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "GPU access is not documented on the pages checked",
      "networking": "None unless the embedder grants WASI capabilities; guests only reach files, directories and hosts they are explicitly given",
      "pricing": "Free and open source",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Embedding a capability-based sandbox for plug-ins, functions and tool code inside your own application",
      "overview": [
        "Wasmtime is \"a standalone runtime for WebAssembly\" from the Bytecode Alliance, written in Rust and licensed Apache-2.0. It compiles Wasm with the Cranelift code generator at runtime or ahead of time, passes the official WebAssembly test suite, implements WASI and the component model (its README example targets wasm32-wasip2), and offers embeddings supported by the Alliance for Rust, C, C++, Python, .NET, Go and Ruby, with community embeddings for Elixir and Perl. The security model is the WebAssembly one: \"all accesses within linear memory are checked to ensure they stay in bounds\", linear memories get a 2 GB guard region, instance memory is zeroed after use, WASI is capability-based so guests reach only files they were given, and Spectre mitigations cover indirect calls and br_table. Hosts can bound execution with deterministic **fuel** or low-overhead **epoch** interruption. The project is fuzzed continuously on Google's OSS-Fuzz. Binaries ship for Linux, macOS and Windows.",
        "Wasmtime is the engine under most of the Wasm platforms on this site: Spin \"takes advantage of the latest developments in the WebAssembly component model and Wasmtime runtime\", the Extism runtime pins wasmtime 49 in its Cargo.toml, and Akamai Functions runs on \"a Wasmtime-based runtime with strong isolation and memory safety\". The adopters file also lists Fastly (every Compute request), Shopify Functions, Microsoft (WASI node pools in AKS), DFINITY, Huawei Cloud, Redpanda, SingleStore and Cosmonic. A new major version ships on the 20th of each month; every twelfth release is an LTS supported for 24 months, others for 2 months. v50.0.1 was released on 29 October 2026. It is a runtime for code compiled to Wasm, not a Linux environment: there is no shell, package manager or process model unless the guest brings one."
      ],
      "notable": [
        "Monthly semver-major releases on the 20th; every twelfth release is a 24-month LTS; v50.0.1 released 29 October 2026",
        "Adopters include Fastly, Shopify, Microsoft (AKS WASI node pools), Akamai, DFINITY, Huawei Cloud and Redpanda; Spin and Extism build on it",
        "Fuzzed 24/7 on Google's OSS-Fuzz; Spectre mitigations and formal-verification work documented"
      ],
      "pros": [
        "You want a memory-safe, capability-based sandbox you can embed in-process in seven supported languages",
        "You need deterministic execution limits (fuel) for untrusted tool or plug-in code",
        "You run on laptops and servers alike, including macOS and Windows, with no hypervisor",
        "You want the same engine that Fastly, Shopify, Spin and Extism rely on"
      ],
      "cons": [
        "Runs only code compiled to WebAssembly; no Linux userland, shell or arbitrary native packages",
        "Networking and filesystem access exist only through capabilities you wire up with WASI",
        "GPU access is not documented",
        "A new major version ships each month and non-LTS releases are supported for 2 months; LTS releases are supported for 24 months"
      ],
      "verdict": "The Bytecode Alliance's embeddable WebAssembly runtime for building a sandbox into your own software; it suits plug-ins and tool code rather than whole-OS workloads.",
      "peers": [
        "wasmer",
        "extism",
        "fermyon-spin",
        "pyodide",
        "webcontainers"
      ],
      "sources": [
        "https://wasmtime.dev",
        "https://github.com/bytecodealliance/wasmtime",
        "https://docs.wasmtime.dev/security.html",
        "https://docs.wasmtime.dev/stability-release.html",
        "https://docs.wasmtime.dev/examples-interrupting-wasm.html",
        "https://github.com/bytecodealliance/wasmtime/blob/main/ADOPTERS.md",
        "https://github.com/bytecodealliance/wasmtime/releases",
        "https://github.com/spinframework/spin",
        "https://github.com/extism/extism/blob/main/runtime/Cargo.toml",
        "https://www.akamai.com/products/akamai-functions"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "wasmer",
      "name": "Wasmer",
      "company": "Wasmer, Inc.",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.ycombinator.com/companies/wasmer"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://wasmer.io",
      "docs": "https://docs.wasmer.io",
      "pricing_url": "https://wasmer.io/pricing",
      "github": "https://github.com/wasmerio/wasmer",
      "tagline": "An MIT-licensed WebAssembly runtime and WASIX ecosystem that runs registry packages such as Python, Bash and PHP in local sandboxes, with Wasmer Edge as the hosted option.",
      "isolation": "WebAssembly sandbox per package; \"No file, network, or environment access, unless explicitly enabled\"; WASIX extends WASI with POSIX-style features",
      "isolation_short": "Wasm sandbox",
      "isolation_class": "wasm",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "CLI",
        "JavaScript/TypeScript",
        "Python",
        "Rust",
        "Swift"
      ],
      "cold_start": "Home page claims \"107× faster\" cold starts in a comparison with Docker and \"51× smaller\" memory per instance (20 MB vs. 1,024 MB); no absolute start time is given",
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Wasmer Edge offers persistent volumes, remote sessions via wasmer ssh and app versioning with rollback; the runtime CLI documents journal files with snapshot triggers (--journal, --snapshot-on), and Wasmer Edge documents InstaBoot snapshots",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "GPU support is not documented",
      "networking": "Disabled unless explicitly enabled; WASIX networking in the runtime, and in the JavaScript SDK Node.js networking is bridged through node:net and node:dns",
      "pricing": "Runtime free and open source. Wasmer Edge: Hobby free; Pro $10/month with overages of $0.50 per million requests, $0.10 per GB bandwidth, $0.10 per active CPU hour, $0.01 per GB-hour memory, $0.05 per GB storage; Enterprise and subhosting custom",
      "free_tier": "Free to self-host; Wasmer Edge Hobby plan free for personal non-commercial projects: 100 apps, 1,000,000 requests, 150 GB bandwidth, 6 active CPU hours and 400 GB-hours memory per month",
      "compliance": null,
      "regions": "Wasmer Edge: 12 global regions on Hobby and Pro; private dedicated locations on Enterprise",
      "best_for": "Running interpreter packages (Python, PHP, Bash) in Wasm sandboxes from JavaScript or Python, locally, in the browser or on Wasmer Edge",
      "overview": [
        "Wasmer is a WebAssembly runtime and CLI, MIT-licensed, that \"runs your apps in fast, secure, and lightweight sandboxes: locally, in the cloud, or in your browser\". The company, backed by Y Combinator, pairs the runtime with a **registry** of ready-to-run packages (Bash, PHP, Python, Node.js, PGlite, SQLite, curl, FFmpeg, Pandoc) and **WASIX**, its superset of WASI. The runtime is \"Secure by default. No file, network, or environment access, unless explicitly enabled\", offers Singlepass, Cranelift and LLVM backends plus browser and V8 integration backends, and installs on macOS, Linux and Windows; wasmer run python/python starts a sandboxed Python. The separate wasmer-sdk \"turns packages from the Wasmer registry into composable sandboxes\" from JavaScript (Node.js 20+ and browsers), Python, Rust and a Swift preview, and is \"currently alpha\".",
        "Wasmer Edge is the hosted side: a platform \"powered by WebAssembly, Wasmer, and WASIX\" that runs stateless HTTP workloads close to users, with volumes, cron jobs, wasmer ssh sessions and rollback, and always-on workloads in development. The Hobby plan is free for non-commercial use and Pro is $10 per month with usage overages; Enterprise adds subhosting, private locations and a 99.95% SLA. Wasmer's home page is headed \"Universal compute. Built for AI\" and its repository describes \"sandboxes for your apps and AI agents\"; an agent-specific sandbox API is not documented on the pages checked. Wasmer v7.5.0 was released on 1 October 2026. Because everything runs as Wasm, native packages that are not in the registry or compiled to WASIX are out of reach."
      ],
      "notable": [
        "Runtime is MIT-licensed; v7.5.0 released 1 October 2026; about 21,000 GitHub stars",
        "Backed by Y Combinator; Wasmer Edge Pro is $10/month, Hobby is free with 12 global regions",
        "wasmer-sdk (JavaScript, Python, Rust, Swift) is labelled alpha"
      ],
      "pros": [
        "You want to run a sandboxed Python or Bash interpreter inside a JavaScript or Python process, including in the browser",
        "You need a runtime that installs on macOS, Windows and Linux without a hypervisor",
        "You want a hosted path (Wasmer Edge) with a free tier and published overage rates"
      ],
      "cons": [
        "Only packages compiled to Wasm/WASIX run; arbitrary native dependencies are unavailable",
        "The sandbox SDK is \"currently alpha\"; its README says error codes and less common capabilities \"may still evolve\"",
        "Wasmer Edge is stateless HTTP by default; TCP/UDP and always-on workloads require contacting the team or are in development",
        "Home-page cold-start and memory figures are relative comparisons (\"107× faster\" than Docker; 20 MB vs. 1,024 MB), not absolute timings"
      ],
      "verdict": "A WebAssembly runtime and SDK for embedding interpreter sandboxes in JavaScript or Python apps; it runs Wasm/WASIX packages rather than a full Linux environment.",
      "peers": [
        "wasmtime",
        "fermyon-spin",
        "pyodide",
        "extism",
        "cloudflare-sandbox"
      ],
      "sources": [
        "https://wasmer.io",
        "https://wasmer.io/pricing",
        "https://docs.wasmer.io",
        "https://docs.wasmer.io/runtime",
        "https://docs.wasmer.io/edge",
        "https://github.com/wasmerio/wasmer",
        "https://github.com/wasmerio/wasmer-sdk",
        "https://github.com/wasmerio/wasmer/releases",
        "https://docs.wasmer.io/runtime/cli/",
        "https://docs.wasmer.io/edge/learn/instaboot/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "fermyon-spin",
      "name": "Spin (Fermyon)",
      "company": "CNCF sandbox project (maintained by Fermyon, part of Akamai)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://spinframework.dev",
      "docs": "https://spinframework.dev/v4/",
      "pricing_url": null,
      "github": "https://github.com/spinframework/spin",
      "tagline": "A CNCF sandbox framework for event-driven WebAssembly component apps on Wasmtime, maintained by Fermyon (now part of Akamai) and hosted as Fermyon Cloud and Akamai Functions.",
      "isolation": "WebAssembly component per handler on the Wasmtime runtime; Akamai Functions describes \"a Wasmtime-based runtime with strong isolation and memory safety\"",
      "isolation_short": "Wasm component",
      "isolation_class": "wasm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "CLI",
        "Rust",
        "JavaScript/TypeScript",
        "Go",
        "Python"
      ],
      "cold_start": "spinframework.dev cites \"Millisecond cold start times\"; the Akamai Functions page says \"zero cold starts\" and \"sub-millisecond startup\"; neither page gives benchmark figures",
      "cold_start_short": "ms (claim)",
      "max_session": "Unlimited self-hosted; Fermyon Cloud Starter caps request handler duration at 30 seconds",
      "max_session_short": "30 s (Fermyon Cloud)",
      "persistence": "Stateless handlers; Fermyon Cloud provides key-value stores and SQLite databases (Starter: 5 stores at 1 GB total, 1 database up to 1 GB)",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": "Not documented for general component code; Spin has a Serverless AI API for LLM inferencing, which can be serviced remotely through a cloud-gpu plugin",
      "networking": "Outbound HTTP from components via the Spin SDKs; Fermyon Cloud Starter allows 500 outbound requests per hour per app",
      "pricing": "Spin: free and open source. Fermyon Cloud: Starter $0/month, Growth $19.38/month, Enterprise custom. Akamai Functions: pricing is not published on the product page; the product must be enabled on an Akamai Control Center contract",
      "free_tier": "Free to self-host; Fermyon Cloud Starter: 5 apps, 100,000 requests per month, 5 GB egress, 1 region",
      "compliance": null,
      "regions": null,
      "best_for": "Serverless HTTP functions compiled to Wasm components, run locally with the Spin CLI or on Fermyon Cloud, SpinKube or Akamai",
      "overview": [
        "Spin is \"an open source framework for building and running fast, secure, and composable cloud microservices with WebAssembly\", licensed Apache-2.0 and \"an open source Cloud Native Computing Foundation sandbox project\". It \"takes advantage of the latest developments in the WebAssembly component model and Wasmtime runtime\": you write an HTTP or Redis-triggered handler in Rust, JavaScript/TypeScript, Go or Python (the framework team supports those four SDKs; Zig and MoonBit are third-party), compile it to a wasm32-wasip2 component, and run it with the spin CLI locally, on your own servers, on Kubernetes through SpinKube, or on a hosted service. Spin v4.3.0 was released on 5 November 2026.",
        "Fermyon Technologies, which maintains Spin, was acquired by Akamai; Akamai's press release of 1 December 2025 says it \"has acquired Fermyon the serverless WebAssembly company\" and that it will continue to support the Spin and SpinKube CNCF projects, and that Fermyon's employees, \"including co-founders Matt Butcher and Radu Matei, will join Akamai's Cloud Technology Group\". Two hosted targets exist: **Fermyon Cloud**, \"currently an open beta service\" with no SLA, whose Starter plan allows 5 apps, 100,000 requests a month and 30-second handlers; and **Akamai Functions** (formerly Fermyon Wasm Functions), a \"globally distributed PaaS\" built on Spin, which must be enabled on an Akamai Control Center contract and is deployed with spin aka deploy. As a function framework, Spin is a building block for tool endpoints rather than a workspace: handlers are short-lived, stateless and limited to languages that compile to Wasm components."
      ],
      "notable": [
        "CNCF sandbox project; Apache-2.0; Spin v4.3.0 released 5 November 2026",
        "Akamai announced its acquisition of Fermyon on 1 December 2025 and committed to continue supporting Spin and SpinKube",
        "Fermyon Cloud is \"currently an open beta service\" without service-level agreements; Akamai Functions must be enabled on an Akamai Control Center contract"
      ],
      "pros": [
        "You want millisecond-start serverless functions in Wasm with a local-first CLI workflow",
        "You run Kubernetes and want the same components deployed via SpinKube",
        "You need a hosted option with a free tier (Fermyon Cloud Starter) or an edge platform (Akamai Functions)"
      ],
      "cons": [
        "Handlers are request-triggered and short-lived (30 s on Fermyon Cloud Starter) rather than long-running workspaces",
        "Only languages with Wasm component toolchains; no native packages",
        "Fermyon Cloud is \"currently an open beta service\" without service-level agreements, Akamai Functions pricing is not published, and Fermyon was acquired by Akamai in December 2025",
        "GPU access for general component code is not documented"
      ],
      "verdict": "A framework for Wasm-isolated, event-driven function endpoints that run locally, on Kubernetes or on hosted services; it runs compiled Wasm components rather than arbitrary native code.",
      "peers": [
        "wasmtime",
        "wasmer",
        "extism",
        "cloudflare-sandbox",
        "fly-machines"
      ],
      "sources": [
        "https://spinframework.dev",
        "https://github.com/spinframework/spin",
        "https://github.com/spinframework/spin/releases",
        "https://www.akamai.com/newsroom/press-release/akamai-announces-acquisition-of-function-as-a-service-company-fermyon",
        "https://www.akamai.com/products/akamai-functions",
        "https://techdocs.akamai.com/akamai-functions/docs/welcome",
        "https://developer.fermyon.com/cloud/index",
        "https://developer.fermyon.com/cloud/faq",
        "https://www.fermyon.com/pricing",
        "https://techdocs.akamai.com/akamai-functions/docs/quickstart",
        "https://spinframework.dev/v4/serverless-ai-api-guide"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "extism",
      "name": "Extism",
      "company": "Dylibso",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://extism.org",
      "docs": "https://extism.org/docs/overview",
      "pricing_url": null,
      "github": "https://github.com/extism/extism",
      "tagline": "A BSD-licensed WebAssembly plug-in system from Dylibso with host SDKs for 15+ languages and manifest-level memory, host and path limits, built on Wasmtime.",
      "isolation": "WebAssembly sandbox per plug-in on Wasmtime (runtime pins wasmtime 49); manifest sets max memory pages, allowed HTTP hosts and allowed filesystem paths; \"secure & host-controlled HTTP without WASI\"",
      "isolation_short": "Wasm plug-in sandbox",
      "isolation_class": "wasm",
      "open_source": true,
      "license": "BSD-3-Clause",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Rust",
        "JavaScript/TypeScript",
        "Go",
        "Python",
        "Ruby",
        ".NET",
        "Java",
        "PHP",
        "Elixir",
        "Haskell",
        "OCaml",
        "Perl",
        "Zig",
        "C",
        "C++"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": "Runtime limiters and timers are listed as features; plug-in calls are bounded by the host",
      "max_session_short": "Host-set",
      "persistence": "None; plug-ins are called with input bytes and return output bytes, with plug-in memory owned by the host",
      "snapshots": false,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": "GPU access is not documented",
      "networking": "Host-controlled HTTP without WASI; allowed_hosts in the manifest (an empty list blocks all hosts, null allows all)",
      "pricing": "Free and open source",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Letting users or agents supply code that runs as a sandboxed plug-in inside your application, in almost any host language",
      "overview": [
        "Extism is \"a lightweight framework for building with WebAssembly\" and a \"plug-in system for everyone\", an open-source product of Dylibso under the BSD-3-Clause licence. The host embeds an Extism SDK, which exists for Rust, JavaScript (web, Node, Deno, Bun), Go, Python, Ruby, .NET, Java, PHP, Elixir, Haskell, OCaml, Perl, Zig, C and C++, and loads plug-ins compiled with a PDK (Rust, JS, Python, Go, Haskell, AssemblyScript, .NET, C, C++, Zig). A plug-in is described by a manifest that doubles as its policy: max_pages caps memory in 64 KiB pages, allowed_hosts limits outbound HTTP (an empty list blocks everything), allowed_paths maps host directories in only when WASI is enabled, and config passes free-form settings. The runtime's Cargo.toml depends on wasmtime 49 with WASI preview 1 support.",
        "The pitch is that you \"determine where in your application some arbitrary code should run, what data that code should be provided, and the data the plug-in should return\", and the sandbox keeps that code from harming the host \"accidentally or maliciously\". That maps directly onto agent tool execution: each tool can be a Wasm plug-in with its own memory and network allow-list, called from whichever language your agent runs in. The trade-off is scope. Plug-ins are functions, not environments: there is no shell, package manager or long-running process, and anything the plug-in needs must be compiled to Wasm. Extism v1.41.0 was released on 9 October 2026."
      ],
      "notable": [
        "BSD-3-Clause; an open-source product from the team at Dylibso; v1.41.0 released 9 October 2026",
        "Host SDKs for 15+ languages and PDKs for 10 plug-in languages",
        "Runtime built on Wasmtime (pinned to major version 49 in Cargo.toml)"
      ],
      "pros": [
        "You want sandboxed, user-supplied or agent-written code called in-process from almost any host language",
        "You need per-plug-in memory caps and HTTP host allow-lists declared in a manifest",
        "You want plug-ins written in several languages to share one interface"
      ],
      "cons": [
        "A plug-in model, not an execution environment: no shell, filesystem by default, or native packages",
        "\"Runtime limiters & timers\" are listed as features, but the manifest page checked documents no timeout field, only memory, hosts, paths and config",
        "GPU access is not documented"
      ],
      "verdict": "A WebAssembly plug-in framework for running user-supplied or agent-written code as sandboxed plug-ins inside your own app; it runs Wasm modules rather than an operating-system environment.",
      "peers": [
        "wasmtime",
        "wasmer",
        "fermyon-spin",
        "pyodide"
      ],
      "sources": [
        "https://extism.org",
        "https://extism.org/docs/overview",
        "https://extism.org/docs/concepts/plug-in-system",
        "https://extism.org/docs/concepts/manifest",
        "https://github.com/extism/extism",
        "https://github.com/extism/extism/blob/main/runtime/Cargo.toml",
        "https://github.com/extism/extism/releases",
        "https://github.com/extism/extism/blob/main/README.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "pyodide",
      "name": "Pyodide",
      "company": "Pyodide (community project)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://pyodide.org",
      "docs": "https://pyodide.org/en/stable/",
      "pricing_url": null,
      "github": "https://github.com/pyodide/pyodide",
      "tagline": "CPython compiled to WebAssembly, with NumPy, pandas and SciPy ported, running Python inside the browser or Node.js sandbox.",
      "isolation": "CPython compiled to WebAssembly/Emscripten, confined by the browser or Node.js WebAssembly sandbox; no sockets, threads or subprocess; HTTP only via the host fetch stack",
      "isolation_short": "CPython in Wasm",
      "isolation_class": "wasm",
      "open_source": true,
      "license": "MPL-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "JavaScript/TypeScript",
        "Python"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Default in-memory filesystem (MEMFS) is lost when the page reloads; IDBFS in browsers or NODEFS in Node.js can be mounted to persist files; snapshotting is not documented on the pages checked",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "GPU access is not documented",
      "networking": "HTTP via the browser fetch stack only (urllib3 and requests work within browser limits); the sockets module imports but does not function; ssl is a stub",
      "pricing": "Free and open source",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Running short Python and data-science snippets in a browser tab or Node.js process without a server-side sandbox",
      "overview": [
        "Pyodide is \"a port of CPython to WebAssembly/Emscripten\", a \"Python distribution for the browser and Node.js\" licensed under the Mozilla Public License 2.0. Michael Droettboom created it at Mozilla in 2018 as part of the Iodide notebook project; it is now \"an independent and community-driven open-source project\" with its own governance document. It loads from a CDN script tag or the pyodide npm package (Node.js 18 or newer since 0.25.0), exposes a foreign function interface so JavaScript and Python can call each other with async/await, and installs packages with micropip: \"Any pure Python package with a wheel available on PyPi is supported\", and many packages with C, C++ and Rust extensions, including NumPy, pandas, SciPy, Matplotlib and scikit-learn, have been ported. Version 314.0.8, released on 13 November 2026, tracks Python 3.14.",
        "The sandbox is the host's WebAssembly sandbox, which is why Pyodide is used as a zero-infrastructure Python interpreter. The constraints follow from that: threading and multiprocessing are \"included but not working\", sockets do not function, ssl is a stub, modules such as fcntl, pwd, resource, curses and tkinter are removed, and HTTP goes through the browser with \"very little or no control over certificates, timeouts, proxies\". Long computations block the main browser thread unless run in a Web Worker. The default filesystem is in-memory and is lost when the page reloads unless a persistent filesystem such as IDBFS or NODEFS is mounted; snapshotting is not documented on the pages checked. Working sockets and native packages without a WebAssembly build are outside what Pyodide documents."
      ],
      "notable": [
        "Created at Mozilla in 2018; independent community project under MPL-2.0",
        "Version 314.0.8 released 13 November 2026, on Python 3.14",
        "Node.js below 18 unsupported since 0.25.0; tested browser minimums Firefox 112, Chrome 112, Safari 16.4"
      ],
      "pros": [
        "You want Python with NumPy, pandas and SciPy running client-side with no server or container",
        "You need a Python interpreter embedded in a Node.js or browser-based agent without leaving the Wasm sandbox",
        "You want MPL-licensed, community-governed software with a long history"
      ],
      "cons": [
        "No threads, sockets, subprocess or TLS; networking is limited to what the browser fetch stack allows",
        "Installs pure-Python wheels and packages built as wasm32/emscripten wheels; other native dependencies will not install",
        "The default in-memory filesystem is lost on page reload unless IDBFS or NODEFS is mounted; snapshots are not documented",
        "The runtime and packages must be downloaded by the page before use; the docs include guidance on optimising size and load time"
      ],
      "verdict": "A CPython distribution compiled to WebAssembly for running Python in a browser or Node.js process; it runs inside the host's WebAssembly sandbox without sockets, threads or subprocesses.",
      "peers": [
        "wasmtime",
        "wasmer",
        "webcontainers",
        "codapi",
        "val-town"
      ],
      "sources": [
        "https://pyodide.org/en/stable/",
        "https://pyodide.org/en/stable/usage/index.html",
        "https://pyodide.org/en/stable/usage/wasm-constraints.html",
        "https://pyodide.org/en/stable/project/about.html",
        "https://pyodide.org/en/stable/project/changelog.html",
        "https://github.com/pyodide/pyodide",
        "https://pyodide.org/en/stable/usage/file-system.html",
        "https://pyodide.org/en/stable/usage/faq.html",
        "https://pyodide.org/en/stable/usage/webworker.html",
        "https://github.com/pyodide/pyodide/releases"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "hyperlight",
      "name": "Hyperlight",
      "company": "Cloud Native Computing Foundation",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://hyperlight.org",
      "docs": "https://github.com/hyperlight-dev/hyperlight/blob/main/docs/README.md",
      "pricing_url": null,
      "github": "https://github.com/hyperlight-dev/hyperlight",
      "tagline": "Apache-2.0 Rust library, now a CNCF sandbox project, that embeds hypervisor-isolated micro-VMs with no guest OS in your application; sibling projects add Python, .NET and Rust sandboxes for agent-written code.",
      "isolation": "Hardware-virtualised micro-VM per sandbox on KVM or MSHV (Linux), WHP (Windows) or Hypervisor.framework (macOS on Apple Silicon), with no guest kernel, OS or devices; guests are purpose-built no_std Rust or C binaries that can only call host functions you register",
      "isolation_short": "Micro-VM, no guest OS",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Rust",
        "Python",
        ".NET"
      ],
      "cold_start": "hyperlight.org: micro-VMs are created in 1–2 ms because there is no guest OS to boot (vendor claim); the README says guest function calls complete in microseconds",
      "cold_start_short": "1–2 ms (claim)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Guest state persists across calls within a sandbox; snapshot() and restore() save and reset VM memory, and since 0.19.0 snapshots form a tree whose layers share memory copy-on-write. Core Hyperlight gives guests no filesystem; hyperlight-sandbox adds a read-only /input and a writable /output directory (by default 5 MiB per file, 20 MiB in total and 20 files)",
      "snapshots": true,
      "pause_resume": null,
      "gpu": false,
      "gpu_detail": "No - hyperlight.org says each guest runs with no guest OS and no devices",
      "networking": "None in core Hyperlight: guests have no network, filesystem or syscalls and reach the host only through registered functions. hyperlight-sandbox keeps network traffic off by default and allows specific domains and HTTP verbs with allow_domain()",
      "pricing": "Free and open source (Apache-2.0)",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Embedding hardware-isolated execution of small functions or agent-written scripts inside your own application, with millisecond sandbox creation",
      "overview": [
        "Hyperlight is \"a lightweight VMM for running untrusted code in micro VMs with minimal overhead\". It was announced in November 2024 by Microsoft's Azure Core Upstream team and is now a Cloud Native Computing Foundation sandbox project, licensed Apache-2.0. You embed it as a Rust library, hand it a guest binary and call typed functions across the VM boundary; there is no kernel or OS in the VM, guests are ELF binaries written in no_std Rust or C, and by default they can only print to the host unless you register more host functions. The README lists support for KVM and MSHV on Linux (x86-64, plus AArch64 on KVM), WHP on Windows (x86-64 and AArch64) and Hypervisor.framework on Apple Silicon macOS. It is explicitly not designed for full Linux workloads that need syscalls, networking or a filesystem.",
        "The agent-facing pieces live in the same GitHub organisation. hyperlight-sandbox (v0.10.0) is a multi-backend framework with Python, .NET and Rust SDKs that runs Python or JavaScript in a Wasm component guest, or JavaScript on HyperlightJS, with schema-validated tool calls, capability-based file access, domain allowlisting and snapshot/restore; its README lists \"Code Mode: Let an agent write a script that calls your tools directly\" as a use case. HyperAgent is \"a sandboxed code-acting AI agent runtime\" built on the GitHub Copilot SDK that writes JavaScript handlers and runs them in Hyperlight micro-VMs; its README warns it is pre-release software, not for production use, and that it does not run on macOS. hyperlight-wasm and hyperlight-unikraft extend the guest options to WebAssembly modules and ordinary Linux programs. The core project is pre-1.0 and warns that upgrades may require code changes."
      ],
      "notable": [
        "Announced by Microsoft's Azure Core Upstream team in November 2024; now a CNCF sandbox project with about 4,800 GitHub stars.",
        "v0.19.0 released 4 November 2026; its release post reports large-snapshot creation dropping from 171 ms to about 1.7 ms (vendor benchmark).",
        "HyperAgent v0.7.0 (July 2026) is labelled pre-release and requires GitHub Copilot access and hardware virtualisation."
      ],
      "pros": [
        "You want hypervisor isolation for small functions or agent-written scripts without booting a guest OS per sandbox.",
        "You need to embed sandboxed execution as a library in a Rust, Python or .NET application.",
        "You want guests that can reach only the host functions and domains you explicitly allow.",
        "You want snapshot and restore to reset a sandbox to a clean state between calls."
      ],
      "cons": [
        "Guests must be purpose-built against the Hyperlight guest library or run through a sibling runtime (Wasm, JavaScript, Unikraft); unmodified Linux programs do not run on core Hyperlight.",
        "Pre-1.0: the API may change between releases.",
        "Lifecycle, networking and scheduling around the sandbox are yours to build; it is a library, not a service.",
        "HyperAgent is pre-release, not for production and not supported on macOS."
      ],
      "verdict": "An embeddable micro-VM library with no guest OS for running untrusted functions and agent-written scripts inside your own application; the agent runtimes built on it are early.",
      "peers": [
        "firecracker",
        "wasmtime",
        "gvisor",
        "libkrun",
        "wasmer",
        "extism"
      ],
      "sources": [
        "https://hyperlight.org",
        "https://hyperlight.org/blog/introducing-hyperlight/",
        "https://hyperlight.org/blog/hyperlight-0190/",
        "https://github.com/hyperlight-dev/hyperlight",
        "https://github.com/hyperlight-dev/hyperlight/blob/main/docs/README.md",
        "https://github.com/hyperlight-dev/hyperlight/releases",
        "https://github.com/hyperlight-dev/hyperlight-sandbox",
        "https://github.com/hyperlight-dev/hyperagent"
      ],
      "verified_on": "2026-11-18"
    },
    {
      "id": "lifo",
      "name": "Lifo",
      "company": "Lifo",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://lifo.sh",
      "docs": "https://lifo.sh/docs",
      "pricing_url": "https://lifo.sh/pricing",
      "github": "https://github.com/lifo-sh/lifo",
      "tagline": "MIT Linux-like OS written in TypeScript (kernel, filesystem, bash-like shell, processes and a Node.js layer) that boots inside a browser tab or Node.js process to run agent-generated apps and shell tools.",
      "isolation": "Language-level: a Unix-like system and Node.js APIs reimplemented in TypeScript, running in the same JavaScript realm as the page or Node.js process that created the box. Boxes are kept apart from each other but not from their host; the docs say \"a box is not a security boundary\". WorkerSandbox moves a box into a Web Worker, optionally on a separate origin, to rely on the browser's isolation",
      "isolation_short": "In-process JS runtime",
      "isolation_class": "isolate",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "JavaScript/TypeScript",
        "CLI"
      ],
      "cold_start": "Docs (vendor measurement with the repo's bench/ in headless Chromium): about 27 ms to a box with node and npm in a browser tab (26 ms to load, under 1 ms to boot) and about 0.5 ms in Node.js",
      "cold_start_short": "~27 ms in a tab (claim)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Files live in memory; in a browser, persist: true saves the filesystem to IndexedDB (one store per origin) and restores it on the next visit, otherwise every box starts clean and a tab's box ends when the tab closes. Snapshots are .tar.gz files of every file plus the working directory and environment, exported and imported from the browser, Node.js or the lifo-sh CLI",
      "snapshots": true,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": "Not documented for programs in a box; the optional lifo-pkg-agent runs a WebLLM model on the visitor's GPU through WebGPU",
      "networking": "No raw sockets: programs make HTTP(S) and WebSocket requests through the host's fetch and WebSocket, under the browser's CORS rules. Servers in a box are reachable from the page, an iframe preview served by a service worker, or a public URL via lifo tunnel, not from other apps on the machine. Cloud boxes (preview) get a public URL and outbound TCP",
      "pricing": "Free under MIT in the browser, Node.js and the terminal; browser execution is never metered. Lifo Cloud (cloud boxes are labelled a preview in the docs): Hobby $0 with $5 of usage a month; Pro $20/month including $20 of usage; Enterprise custom. Rates: Lite box $0.005 per awake hour; Full box (coming soon) $0.08, $0.16 or $0.32 per hour for S, M or L; headless browser $0.05 per browser hour; tools such as ffmpeg $0.10 per CPU hour; storage $0.15 per GB-month; data out $0.06 per GB",
      "free_tier": "Free and unmetered in the browser or Node.js; Lifo Cloud Hobby includes $5 of usage a month with no card",
      "compliance": null,
      "regions": null,
      "best_for": "Running agent-generated web apps, previews and a bash tool for agents in the user's own browser tab or a Node.js process, where the code is trusted or the browser provides the isolation",
      "overview": [
        "Lifo is \"a small Linux-like operating system written in TypeScript\": a kernel, a POSIX-like virtual filesystem, a bash-like shell with 60+ commands, processes, ports and a Node.js compatibility layer, with no VM or emulator underneath. The same @lifo-sh/core package runs in a browser tab, in a Node.js process and from the lifo-sh CLI, and a box moves between them as a snapshot. Real npm, Vite, Express and Expo run in it, Postgres comes through PGlite, git through isomorphic-git, and python3 is CPython on WebAssembly. It cannot run native Linux binaries or native npm addons, and long computations share one thread with the page. The website pitches it as \"Your agent builds it. Lifo runs it.\" and the pricing page lists \"AI agent sandboxes\"; the docs describe giving a model one bash tool backed by box.commands.run() and an in-tab agent package that runs WebLLM.",
        "Isolation is the main caveat, and the vendor states it plainly: programs in a box are kept apart from each other, but their JavaScript runs in the same realm as the page or Node.js process that created the box, so it can reach the page's DOM, cookies and storage or the real process, filesystem and network. The docs say \"a box is not a security boundary\" and recommend WorkerSandbox, ideally served from a different origin inside a sandboxed iframe, for code you do not trust, or a cloud sandbox when you need real isolation. The runtime is MIT; Lifo Cloud, whose servers are closed source, has published pricing, but the docs label cloud boxes a preview that is not yet available on lifo.sh. The project is pre-1.0 (0.10.x) and is built by the team behind RapidNative, which the website names as its production user."
      ],
      "notable": [
        "@lifo-sh/core and lifo-sh 0.10.36 published to npm on 17 November 2026; 565 GitHub stars at verification.",
        "Vendor benchmark: ~27 ms cold start in a tab and ~0.3 MB gzipped runtime, against ~6.6 s measured for WebContainers in the same browser.",
        "Docs: cloud boxes are in development and lifo cloud reports the cloud \"isn't available yet\", although pricing is published."
      ],
      "pros": [
        "You want agent-generated web apps to run in the user's own tab with no per-session server cost.",
        "You need a disposable shell, filesystem, node and npm for an agent tool inside a Node.js process.",
        "You want the same box in the browser, Node.js and the terminal, moved between them as a snapshot.",
        "You want an MIT runtime with no license fees or session caps."
      ],
      "cons": [
        "Not a security boundary: code in a box can reach the page or Node.js process that runs it unless you add a Web Worker and a separate origin.",
        "No native Linux binaries or native npm addons; Node.js is a compatibility layer and the shell is bash-like, not bash.",
        "Processes share one JavaScript thread, so long computations block other programs and the page.",
        "Lifo Cloud is priced but labelled a preview and not yet available; pre-1.0 releases."
      ],
      "verdict": "A TypeScript reimplementation of a Unix userland for running agent-generated apps in a browser tab or Node.js process; its isolation is JavaScript-level, not a security boundary, so untrusted code needs a worker on a separate origin or a real sandbox.",
      "peers": [
        "webcontainers",
        "just-bash",
        "pyodide",
        "langchain-sandbox",
        "wasmer",
        "e2b"
      ],
      "sources": [
        "https://lifo.sh",
        "https://lifo.sh/pricing",
        "https://lifo.sh/docs",
        "https://lifo.sh/docs/understand/why",
        "https://lifo.sh/docs/understand/security",
        "https://lifo.sh/docs/understand/limitations",
        "https://lifo.sh/docs/understand/performance",
        "https://lifo.sh/docs/understand/tab-and-cloud",
        "https://lifo.sh/docs/build/files",
        "https://lifo.sh/docs/build/agents",
        "https://lifo.sh/docs/agents",
        "https://lifo.sh/docs/cloud/boxes",
        "https://lifo.sh/docs/reference/cli",
        "https://lifo.sh/license",
        "https://github.com/lifo-sh/lifo",
        "https://registry.npmjs.org/@lifo-sh/core"
      ],
      "verified_on": "2026-11-18"
    },
    {
      "id": "smolvm",
      "name": "Celesto (SmolVM)",
      "company": "Celesto AI",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://celesto.ai",
      "docs": "https://celesto.ai/docs/",
      "pricing_url": "https://celesto.ai/pricing",
      "github": "https://github.com/CelestoAI/celesto",
      "tagline": "Open-source microVM \"computers\" for agents from Celesto AI, formerly published as SmolVM, with Firecracker on Linux and QEMU on macOS, plus a hosted cloud option.",
      "isolation": "MicroVM per computer with hardware-level separation from the host: Firecracker on Linux (KVM), QEMU on macOS; the PyPI listing also names libkrun as a backend",
      "isolation_short": "Firecracker / QEMU microVM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "CLI",
        "MCP server"
      ],
      "cold_start": "README: a local computer \"starts in about 500 ms\"; the website says hosted agent sandboxes \"start in under 2 seconds\"",
      "cold_start_short": "~500 ms (claim)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Snapshots \"pause and restore memory, disk, and active processes\"; persistent state survives across sessions; in the cloud, stopped computers keep files, packages and workspace state, and stopped root disks are kept free for 30 days",
      "snapshots": true,
      "pause_resume": true,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Restricted-destination network controls are documented for Linux only (Firecracker, or QEMU with TAP); QEMU slirp supports open and off modes on macOS and Linux; the QEMU backend uses user-mode (slirp) networking with port forwarding; cloud computers can expose public ports",
      "pricing": "Cloud computers billed per hour of running time, prepaid with no monthly fee: Micro 1 vCPU/2 GB $0.015/h, Small 1 vCPU/4 GB $0.020/h, Standard 2 vCPU/4 GB $0.030/h, Memory 2 vCPU/8 GB $0.040/h, Large 4 vCPU/16 GB $0.080/h, XL 8 vCPU/32 GB $0.160/h; root disk included; paid accounts start with a default limit of 20 running sandboxes. Local runtime is free",
      "free_tier": "One-time $5 Micro runtime credit, no card, expires 30 days after activation (two retained Micro sandboxes); free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Running agent computers as microVMs on your own Linux or Apple Silicon machine, with the same API available as a hosted cloud computer",
      "overview": [
        "Celesto is an Apache-2.0 runtime from Celesto AI that gives an agent a \"secure and persistent computer\": a microVM booted with Firecracker on Linux or QEMU on macOS (Apple Silicon), driven from a Python SDK (`from celesto import Computer`), a TypeScript SDK, a `celesto` CLI and an MCP server. The project was published as **SmolVM** (the `smolvm` package on PyPI, 0.0.36 in September 2026, classified \"3 - Alpha\") and the GitHub repository CelestoAI/SmolVM now resolves to CelestoAI/celesto. Computers can run a browser session with a CDP endpoint, a desktop with GUI apps, and Windows guests on a Linux host with KVM.",
        "State is a documented focus: snapshots capture memory, disk and processes, and persistent state survives across sessions. The same `Computer` API runs locally or against Celesto's cloud with `provider=\"cloud\"`, billed per hour from $0.015 for a 1 vCPU/2 GB Micro computer. Maturity labels in the sources: \"3 - Alpha\" on PyPI, \"alpha\" for the TypeScript SDK on the smolvm PyPI page, \"preview\" for the macOS desktop; the docs say Windows guest support \"is rolling out in phases\". The company also sells GitHub Actions runners on the same infrastructure."
      ],
      "notable": [
        "PyPI package celesto 0.1.13 (5 November 2026) carries development status \"3 - Alpha\"; the earlier smolvm 0.0.36 page labels the TypeScript SDK alpha",
        "README claims a computer \"starts in about 500 ms\"; the website claims hosted sandboxes \"start in under 2 seconds\"",
        "Hosted computers from $0.015 per hour (Micro, 1 vCPU/2 GB) with a one-time $5 credit"
      ],
      "pros": [
        "You want a microVM runtime that works on an Apple Silicon laptop (QEMU) and on Linux (Firecracker) behind one API",
        "You need snapshots that capture memory and running processes, not just a disk image",
        "You want a cheap hosted fallback with the same SDK when a local machine is not available",
        "You are building computer-use agents and want a browser or desktop inside the VM"
      ],
      "cons": [
        "Alpha-labelled on PyPI; renamed from SmolVM to Celesto, with the smolvm command and Python import removed",
        "Restricted-destination network controls are Linux only; QEMU on macOS supports only open and off modes",
        "Compliance, regions and session limits are not documented"
      ],
      "verdict": "An alpha-labelled open-source microVM runtime that runs locally on macOS or Linux, with a per-hour hosted cloud for teams that want the same API managed.",
      "peers": [
        "boxlite",
        "microsandbox",
        "arrakis",
        "firecracker",
        "e2b",
        "agents-sandbox"
      ],
      "sources": [
        "https://celesto.ai",
        "https://celesto.ai/docs/",
        "https://celesto.ai/docs/cloud/overview.md",
        "https://celesto.ai/pricing",
        "https://github.com/CelestoAI/celesto",
        "https://pypi.org/project/smolvm/",
        "https://celesto.ai/docs/celesto/installation.md",
        "https://celesto.ai/docs/celesto/concepts/backends.md",
        "https://celesto.ai/docs/celesto/features/network-controls.md",
        "https://pypi.org/project/celesto/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "boxlite",
      "name": "BoxLite",
      "company": "BoxLite",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://boxlite.ai",
      "docs": "https://docs.boxlite.ai",
      "pricing_url": "https://boxlite.ai",
      "github": "https://github.com/boxlite-ai/boxlite",
      "tagline": "Embeddable, daemonless microVM library that runs OCI images in hardware-isolated boxes on a laptop or server, with an optional managed cloud.",
      "isolation": "Hardware-virtualised microVM with its own guest kernel (KVM on Linux, Hypervisor.framework on macOS), plus a jailer around the host-side shim: seccomp allowlist, mount/PID namespaces and privilege drop on Linux, Seatbelt on macOS",
      "isolation_short": "microVM + jailer",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "Go",
        "Rust",
        "C",
        "CLI",
        "REST API"
      ],
      "cold_start": "Docs: boxes \"start in about a second\"; the README gives no boot-time figure; the FAQ lists \"~1-2s\" startup with a cached image",
      "cold_start_short": "~1 s (claim)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Boxes persist across stop/restart and \"resume across turns\"; per-box QCOW2 copy-on-write disk, volume mounts (ro/rw), clone, and export/import as .boxlite archives; detached boxes outlive the parent process",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Outbound internet with an egress allow-list (allow_net), local TCP port forwarding, local/remote tunnels; secrets are injected so real values never enter the VM; user-mode network stack with a fixed subnet",
      "pricing": "Open-source runtime is free. BoxLite Cloud: pay-as-you-go from $0.0666 per hour for the smallest box; Starter $19/month, Pro $149/month, Max $499/month, Enterprise custom; boxes up to 4 vCPU and 32 GiB memory",
      "free_tier": "$100 free credits on BoxLite Cloud; free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Embedding a microVM sandbox directly in your own process on a laptop or server without running a daemon",
      "overview": [
        "BoxLite is an Apache-2.0 runtime whose \"Box\" is a hardware-isolated microVM that runs any OCI image. It is a library first: Python, Node.js, Go, Rust and C bindings call a Rust core that boots a VM on KVM (Linux, including WSL2) or Hypervisor.framework (Apple Silicon macOS) with no daemon, and an optional `boxlite serve` mode exposes a REST API with WebSocket exec. The security docs describe two layers: the VM boundary with a separate guest kernel, and a jailer around the host-side shim (seccomp, namespaces and an unprivileged uid on Linux; Seatbelt on macOS), and they advise \"For running untrusted code in production, prefer Linux\".",
        "Boxes keep a QCOW2 copy-on-write disk, persist across stop/restart, can be cloned and exported as archives, and support volume mounts, PTYs, per-command timeouts and an egress allow-list. BoxLite Cloud sells the same API as a managed service from $0.0666 per hour with $100 of starting credit. Neither the README nor the site uses an alpha/beta label; macOS on Intel is listed as \"Not supported\"."
      ],
      "notable": [
        "Docs claim boxes \"start in about a second\"",
        "BoxLite Cloud pay-as-you-go from $0.0666 per hour; plans at $19, $149 and $499 per month",
        "Supports Linux x86_64/ARM64 (KVM), Apple Silicon macOS 12+ and Windows via WSL2; Intel macOS \"Not supported\""
      ],
      "pros": [
        "You want a microVM sandbox as an in-process library with bindings in five languages",
        "You need boxes that survive restarts and can be cloned or exported",
        "You run on Apple Silicon laptops as well as Linux servers",
        "You want a managed cloud with the same API as the local runtime"
      ],
      "cons": [
        "Requires hardware virtualisation; the docs say it is \"Not available in most CI runners, nested VMs, or containers\"",
        "macOS jailer has no privilege drop, cgroups or network namespace; the docs prefer Linux for untrusted code",
        "GPU support, compliance certifications and regions are not documented; local snapshots capture disk state and must be restored on a stopped box"
      ],
      "verdict": "A practical choice when you want microVM isolation embedded in your own process rather than behind a service.",
      "peers": [
        "smolvm",
        "microsandbox",
        "arrakis",
        "libkrun",
        "firecracker",
        "e2b"
      ],
      "sources": [
        "https://github.com/boxlite-ai/boxlite",
        "https://boxlite.ai",
        "https://docs.boxlite.ai",
        "https://docs.boxlite.ai/architecture/index.md",
        "https://docs.boxlite.ai/architecture/security.md",
        "https://docs.boxlite.ai/getting-started/installation.md",
        "https://docs.boxlite.ai/manage-sandbox/snapshots.md",
        "https://docs.boxlite.ai/faq.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "arrakis",
      "name": "Arrakis",
      "company": "Arrakis (abshkbh)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/abshkbh/arrakis",
      "docs": "https://github.com/abshkbh/arrakis#readme",
      "pricing_url": null,
      "github": "https://github.com/abshkbh/arrakis",
      "tagline": "Self-hosted microVM sandbox for agent code execution and computer use, with snapshot-and-restore and a VNC-accessible GUI in every VM.",
      "isolation": "MicroVM per sandbox using Cloud Hypervisor as the VMM on Linux/KVM; overlayfs protects each sandbox root filesystem",
      "isolation_short": "Cloud Hypervisor microVM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "AGPL-3.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Python",
        "REST API",
        "CLI",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Snapshot-and-restore \"out of the box\" so an agent can return to an earlier state; a restored VM currently reuses the original IP, so the original must be stopped or destroyed first on the same host",
      "snapshots": true,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Each sandbox gets a tap device on a Linux bridge on the host; SSH access and automatic port forwarding; VNC exposed through port forwarding with a noVNC proxy",
      "pricing": "Free to self-host under AGPL-3.0; commercial licensing by email",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Self-hosting a GUI-capable microVM sandbox with snapshot/restore for computer-use and code-execution agents",
      "overview": [
        "Arrakis is \"a fully customizable and self-hosted sandboxing solution for AI agent code execution and computer use\" maintained by GitHub user abshkbh. Each sandbox is a microVM run by Cloud Hypervisor on a Linux host with /dev/kvm (the README includes a GCE setup guide); a Go REST server (`arrakis-restserver`) manages VMs, and you drive it from the `py-arrakis` Python SDK, the `arrakis-client` CLI, the REST API or a separate MCP server for Claude Desktop, Cursor and Windsurf.",
        "Every VM starts a VNC server at boot with Chrome preinstalled, so computer-use agents get a desktop as well as a shell, and snapshot-and-restore lets an agent back out of a bad state. Licensing is AGPL-3.0 with commercial licences offered by email. The README carries no stability label; the Releases page lists numbered builds, the latest being release-36 (2 June 2025); a contribution guide is \"going to be available soon\"."
      ],
      "notable": [
        "AGPL-3.0 licence with commercial licensing offered by email",
        "Snapshot-and-restore and a VNC/noVNC desktop with Chrome in every sandbox",
        "Linux-only host requirement: Cloud Hypervisor needs /dev/kvm"
      ],
      "pros": [
        "You want a self-hosted microVM sandbox that includes a desktop for computer-use agents",
        "You need snapshot/restore so agents can roll back",
        "You want an MCP server so desktop agent clients can use the sandbox directly"
      ],
      "cons": [
        "Linux/KVM hosts only; no macOS or Windows support documented",
        "AGPL-3.0 licence; commercial licensing is by email",
        "Maturity label, boot-time figures and resource limits are not documented in the README"
      ],
      "verdict": "An AGPL-3.0 microVM sandbox for self-hosters on Linux that bundles a VNC desktop and snapshot/restore.",
      "peers": [
        "smolvm",
        "boxlite",
        "microsandbox",
        "firecracker",
        "cua",
        "e2b"
      ],
      "sources": [
        "https://github.com/abshkbh/arrakis",
        "https://github.com/abshkbh/arrakis/releases"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "langchain-sandbox",
      "name": "LangChain Sandbox",
      "company": "LangChain",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://www.langchain.com/about"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/langchain-ai/langchain-sandbox",
      "docs": "https://github.com/langchain-ai/langchain-sandbox#readme",
      "pricing_url": null,
      "github": "https://github.com/langchain-ai/langchain-sandbox",
      "tagline": "Archived Python package that runs untrusted Python in Pyodide (WebAssembly) inside a Deno subprocess; LangChain says it is no longer maintained.",
      "isolation": "Pyodide (CPython compiled to WebAssembly) hosted by Deno in a subprocess; isolation depends on the Deno permissions granted, and broad grants can expose host resources",
      "isolation_short": "Pyodide in Deno",
      "isolation_class": "wasm",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Python"
      ],
      "cold_start": "README: \"There is a few seconds of latency when starting the sandbox per run\"",
      "cold_start_short": "a few seconds",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Optional stateful mode returns session_bytes and session_metadata that can be passed to later execute() calls; no filesystem access, files written in the sandbox cannot be retrieved",
      "snapshots": false,
      "pause_resume": false,
      "gpu": false,
      "gpu_detail": "Not documented; the runtime is Pyodide (WebAssembly) under Deno",
      "networking": "allow_net limits which network requests sandboxed code can make (the README warns about SSRF); examples enable it so Pyodide can install packages; use httpx rather than requests",
      "pricing": "Free (MIT)",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Reference only: a lightweight Pyodide-in-Deno Python sandbox pattern, now archived",
      "overview": [
        "LangChain Sandbox is a Python package that executes model-generated Python inside Pyodide, the WebAssembly build of CPython, hosted by a Deno process. The Python side (`sandbox-py`) spawns a TypeScript module (`pyodide-sandbox-js`) as a subprocess, so Deno must be installed; Deno permissions such as `allow_net` define the boundary. There is no filesystem bridge, and stateful runs are carried as pickled session bytes between calls.",
        "The repository is archived and read-only. The README states \"This package is no longer maintained\" and \"We do not recommend using `langchain-sandbox` for any production use cases\". It remains a compact, MIT-licensed example of the Pyodide-in-Deno approach."
      ],
      "notable": [
        "README: \"This package is no longer maintained\"; repository archived and read-only",
        "README: \"There is a few seconds of latency when starting the sandbox per run\""
      ],
      "pros": [
        "You want a minimal, dependency-light example of running Python in WebAssembly from Python",
        "You need per-call isolation with no persistent filesystem and an explicit network allow-list"
      ],
      "cons": [
        "No longer maintained; LangChain advises against production use",
        "No filesystem access; Python only; a few seconds of startup latency per run",
        "Security guarantees depend on how you configure the sandbox and Deno permissions"
      ],
      "verdict": "An archived, no-longer-maintained package for running Python in Pyodide under Deno; its README advises against production use and points to sandbox APIs or LLM provider APIs instead.",
      "peers": [
        "pyodide",
        "deno-sandbox",
        "llm-sandbox",
        "just-bash",
        "wasmtime"
      ],
      "sources": [
        "https://github.com/langchain-ai/langchain-sandbox"
      ],
      "status": "discontinued",
      "status_label": "archived",
      "status_note": "LangChain archived the langchain-sandbox repository on January 14, 2026 (it is now read-only), and the README says the package is no longer maintained.",
      "verified_on": "2026-11-14"
    },
    {
      "id": "llm-sandbox",
      "name": "LLM Sandbox",
      "company": "LLM Sandbox (vndee)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://vndee.github.io/llm-sandbox/",
      "docs": "https://vndee.github.io/llm-sandbox/",
      "pricing_url": null,
      "github": "https://github.com/vndee/llm-sandbox",
      "tagline": "Python library that runs model-generated code through one API on Docker, Podman or Kubernetes, with Bubblewrap/nsjail local backends and an MCP server.",
      "isolation": "Container per session on Docker, Podman or Kubernetes; optional Bubblewrap or NsJail backends on Linux that share the host kernel and are \"not a VM boundary\"; experimental Windows adapters (Windows MXC, WSL2 Bubblewrap)",
      "isolation_short": "Docker/Podman/K8s container",
      "isolation_class": "container",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Python",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Interactive Python sessions keep state between runs; container pooling for reuse; file copy to and from the sandbox; artifact extraction of plots. Local Linux backends do not support images, pooling or interactive sessions",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Network isolation and resource limits (CPU, memory, time) per session; security policies are advisory (is_safe() returns a verdict without blocking run())",
      "pricing": "Free (MIT)",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Python applications that need to run model-generated code in whatever container runtime they already have",
      "overview": [
        "LLM Sandbox \"runs Large Language Model (LLM) generated code through a common API\" over several backends: Docker, Podman and Kubernetes containers, and on Linux the Bubblewrap and NsJail OS sandboxes (which the README is explicit share the host kernel). It supports Python, JavaScript/Node.js, Java, C++, Go and R, extracts plots as artifacts, pools containers, keeps interactive Python sessions, and ships an MCP server so clients such as Claude Desktop can execute code in it. Examples cover eleven agent frameworks including LangChain and LlamaIndex.",
        "It is a library rather than a service, so limits, lifetime and persistence are whatever the backend gives you. Security policies are advisory: `is_safe()` reports a verdict but does not stop `run()`. The Windows adapters are labelled \"experimental adapters with different prerequisites and API limits\". MIT-licensed; `pip install llm-sandbox` with extras per backend."
      ],
      "notable": [
        "Backends: Docker, Podman, Kubernetes, Bubblewrap and NsJail (Linux), plus experimental Windows MXC and WSL2 adapters",
        "Six languages (Python, JavaScript, Java, C++, Go, R) and an MCP server",
        "README: \"Security policies are advisory\""
      ],
      "pros": [
        "You want one Python API over Docker, Podman and Kubernetes",
        "You need plots and files extracted from the sandbox for data-analysis agents",
        "You run a desktop MCP client and want it to execute code in a container"
      ],
      "cons": [
        "Isolation is provided by the container runtime you choose; the README says the local Linux backends \"share the host kernel and are not a VM boundary\"",
        "Security policies do not block execution on their own",
        "Startup figures, GPU support and a hosted option are not documented"
      ],
      "verdict": "A convenient backend-agnostic wrapper for teams that already run containers and want a Python-first code-execution layer.",
      "peers": [
        "langchain-sandbox",
        "swe-rex",
        "openhands-runtime",
        "docker-sandboxes",
        "judge0",
        "bubblewrap"
      ],
      "sources": [
        "https://github.com/vndee/llm-sandbox",
        "https://vndee.github.io/llm-sandbox/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "openhands-runtime",
      "name": "OpenHands Runtime",
      "company": "OpenHands",
      "hq": {
        "entity": "All Hands AI",
        "city": "Cambridge",
        "region": "Massachusetts",
        "country": "US",
        "source": "https://openhands.dev/privacy"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://docs.openhands.dev",
      "docs": "https://docs.openhands.dev/sdk/guides/agent-server/overview",
      "pricing_url": null,
      "github": "https://github.com/OpenHands/software-agent-sdk",
      "tagline": "The sandbox layer of the OpenHands agent SDK: an agent server that runs in a Docker container, in-process locally, or on the hosted OpenHands Runtime API.",
      "isolation": "Docker container running the agent-server image (DockerWorkspace), with the agent proxying commands and file transfer over HTTP/WebSocket; a local workspace runs in the same process rather than in a sandbox; the hosted Runtime API runs the image in a \"sandboxed environment\" whose mechanism is not named",
      "isolation_short": "Docker container",
      "isolation_class": "container",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Docker workspaces are \"automatically stopped and cleaned up\" when the context exits; persistence beyond a session is not documented on the runtime pages",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Agent server exposed on a host port (examples use 8010); extra_ports exposes VS Code Web (port+1) and VNC (port+2); client talks over HTTP and WebSocket",
      "pricing": "Free to self-host (MIT). The hosted Runtime API requires a RUNTIME_API_KEY; pricing is not published on the runtime docs",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Running OpenHands agents in disposable Docker sandboxes, locally or in CI, with the same server reachable through a hosted runtime for large evaluations",
      "overview": [
        "The OpenHands software agent SDK (\"Python, TypeScript, and REST APIs for building agents that work with code\") separates the agent from where it runs. The runtime layer is the **agent server**: `DockerWorkspace` pulls a pre-built image such as `ghcr.io/openhands/agent-server:latest-python`, starts the container, waits for the server and tears it down afterwards; `DockerDevWorkspace` builds from a custom base image; a local workspace runs in the same process rather than in a sandbox; and `APIRemoteWorkspace` connects to the hosted OpenHands Runtime API. The SDK README also mentions ephemeral workspaces \"in Docker or Kubernetes\".",
        "The Docker image bundles the execution tools, optional VS Code Web and VNC for browser tools. The docs warn that the Runtime API is \"designed primarily for benchmark evaluation at scale\" and point production users to OpenHands Cloud Workspace with \"fully managed sandbox environments with SaaS credential support\". The runtime and SDK are MIT-licensed; no maturity label appears on the runtime pages. This record covers the runtime layer only, not the OpenHands agent or app."
      ],
      "notable": [
        "Agent server images published at ghcr.io/openhands/agent-server; exposes a REST/WebSocket API",
        "Hosted Runtime API is documented as \"designed primarily for benchmark evaluation at scale\"",
        "SDK licensed MIT and powers the OpenHands CLI and OpenHands Cloud"
      ],
      "pros": [
        "You run OpenHands agents and want their execution isolated in disposable containers",
        "You need a VNC-visible browser and VS Code inside the sandbox",
        "You run large benchmark sweeps and want a hosted runtime for the same image"
      ],
      "cons": [
        "The documented sandbox is a Docker container; a microVM option is not documented",
        "Persistence, limits and hosted pricing are not documented on the runtime pages",
        "Part of the OpenHands SDK rather than a standalone general-purpose sandbox API"
      ],
      "verdict": "The sandbox layer of the OpenHands SDK, suited to teams already building agents on that SDK who want them run in Docker containers or on the hosted runtime.",
      "peers": [
        "swe-rex",
        "llm-sandbox",
        "docker-sandboxes",
        "daytona",
        "e2b",
        "gke-agent-sandbox"
      ],
      "sources": [
        "https://docs.openhands.dev/sdk/guides/agent-server/overview",
        "https://docs.openhands.dev/sdk/guides/agent-server/docker-sandbox",
        "https://docs.openhands.dev/sdk/guides/agent-server/api-sandbox",
        "https://github.com/OpenHands/software-agent-sdk",
        "https://docs.openhands.dev/sdk/guides/agent-server/local-server"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "swe-rex",
      "name": "SWE-ReX",
      "company": "SWE-agent",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://swe-rex.com/",
      "docs": "https://swe-rex.com/latest/",
      "pricing_url": null,
      "github": "https://github.com/SWE-agent/SWE-ReX",
      "tagline": "Runtime interface from the SWE-agent team for running agent shell sessions locally or in Docker, Modal and AWS Fargate sandboxes, built for massively parallel runs.",
      "isolation": "Depends on the deployment: Docker or Podman container (DockerDeployment), Modal or AWS Fargate (ModalDeployment, FargateDeployment), a remote host, or no isolation with LocalDeployment; a FastAPI server inside the environment relays commands",
      "isolation_short": "Docker / Modal / Fargate",
      "isolation_class": "container",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "Python",
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Multiple interactive shell sessions stay open for the life of the deployment; file read/write through the runtime; no snapshot or persistence beyond the deployment documented",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "The container runtime server listens on port 8000, mapped to a free host port; other deployments expose the server over HTTP; network policy is not documented",
      "pricing": "Free (MIT); Modal and Fargate backends are billed by those providers",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Running many SWE-agent-style shell sessions in parallel across local, Docker, Modal or Fargate backends",
      "overview": [
        "SWE-ReX is \"a runtime interface for interacting with sandboxed shell environments\" that came out of the SWE-agent project. A `Deployment` starts the environment (Docker/Podman, Modal, AWS Fargate, a remote host or the local machine), a FastAPI server inside it executes commands, and the Python client gets a `RemoteRuntime` with the same interface as the local runtime. It recognises when commands finish, returns output and exit codes, supports interactive tools such as ipython and gdb, and keeps several shell sessions open at once.",
        "The README claims \"Running 100 agents in parallel? No problem either!\" and demos 30 SWE-bench instances at once; Daytona support is marked \"(WIP)\". MIT-licensed, `pip install swe-rex` with extras for Modal and Fargate; no version or release status is stated in the docs."
      ],
      "notable": [
        "Backends: local, Docker/Podman, Modal, AWS Fargate, remote; Daytona marked \"(WIP)\"",
        "README claims \"fast, massively parallel agent runs\" and shows 30 SWE-bench instances in parallel",
        "MIT licence; FastAPI server inside each environment"
      ],
      "pros": [
        "You run SWE-agent or similar and want to swap execution backends without changing agent code",
        "You need many parallel shell sessions for benchmark evaluation",
        "You want to run on Modal or Fargate without writing provider-specific glue"
      ],
      "cons": [
        "Isolation is whatever the chosen backend provides; LocalDeployment has none",
        "Snapshots, persistence and network policy are not documented",
        "Release status is not stated; the docs describe it as built for developing and evaluating agents"
      ],
      "verdict": "A backend-agnostic shell runtime from the SWE-agent project for running many coding-agent sessions in parallel on your own Docker, Modal or Fargate backends.",
      "peers": [
        "openhands-runtime",
        "llm-sandbox",
        "modal",
        "daytona",
        "docker-sandboxes",
        "prime-intellect-sandboxes"
      ],
      "sources": [
        "https://github.com/SWE-agent/SWE-ReX",
        "https://swe-rex.com/latest/",
        "https://swe-rex.com/latest/architecture/",
        "https://swe-rex.com/latest/api/deployments/docker/",
        "https://swe-rex.com/latest/usage/"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "nono",
      "name": "nono",
      "company": "nolabs",
      "hq": null,
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://nono.sh",
      "docs": "https://nono.sh/docs",
      "pricing_url": null,
      "github": "https://github.com/nolabs-ai/nono",
      "tagline": "Kernel-enforced OS-native sandbox for coding agents that runs each tool call in its own ephemeral sandbox, brokers credentials and keeps a tamper-evident audit chain.",
      "isolation": "OS-level sandboxing via Landlock on Linux and Seatbelt on macOS, with opt-in user, network, IPC and UTS namespaces on Linux; a supervisor launches each tool call in a short-lived sandbox with only the capabilities it needs",
      "isolation_short": "Landlock / Seatbelt",
      "isolation_class": "os",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": false,
      "byoc": false,
      "sdks": [
        "CLI",
        "Rust",
        "Python",
        "JavaScript/TypeScript",
        "Go"
      ],
      "cold_start": "Website: \"zero setup and zero latency\" (claim); no figures given",
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Runs on the host filesystem within allow-listed paths; rollback restores tracked files from verified snapshots",
      "snapshots": true,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Domain filtering and API rules by HTTP method and path through a proxy that injects credentials at the boundary; the real credential \"never enters the child\"; block_network() in the SDKs",
      "pricing": "Free (Apache-2.0); a \"nolabs Enterprise\" offering is mentioned without pricing",
      "free_tier": "Free",
      "compliance": null,
      "regions": null,
      "best_for": "Running Claude Code, Codex and other coding agents on your own machine with kernel-enforced file and network allow-lists and brokered credentials",
      "overview": [
        "nono, from nolabs (\"Created by the team behind Sigstore\"), wraps an agent process in an OS-enforced sandbox: Landlock on Linux and Seatbelt on macOS, with optional Linux namespaces, and runs on WSL2. You start an agent with a signed profile from the registry (`nono run --profile claude-code -- claude`); supported agents include Claude Code, Codex, OpenCode, Goose, GitHub Copilot and others. The distinctive part is tool brokering: \"Every tool call gets an ephemeral micro toolbox\", where a supervisor verifies the executable and arguments and launches a short-lived sandbox, while credentials stay with the supervisor and are injected by a proxy.",
        "Sessions produce a hash-chained audit log sealed with a Merkle root, tracked files can be rolled back from verified snapshots, and skills, hooks and instruction files can be verified with Sigstore. The core is a Rust library with Python, TypeScript and Go SDKs. The README says \"In the lead-up to a 1.0 release, APIs are stabilizing\"; no alpha or beta label is used."
      ],
      "notable": [
        "README: \"In the lead-up to a 1.0 release, APIs are stabilizing\"",
        "Enforcement via Landlock (Linux) and Seatbelt (macOS); macOS, Linux and Windows (WSL2) supported",
        "Signed agent profiles distributed through registry.nono.sh"
      ],
      "pros": [
        "You want kernel-enforced allow-lists around an existing coding agent without containers or VMs",
        "You need credentials kept out of the agent process and injected per request",
        "You want a tamper-evident audit trail and file rollback for agent sessions",
        "You are on macOS or Linux and want zero-setup enforcement"
      ],
      "cons": [
        "OS-level sandboxing (Landlock, Seatbelt) on the host kernel rather than a VM",
        "Pre-1.0 with APIs still stabilising",
        "Windows is supported via WSL2; a hosted option is not documented"
      ],
      "verdict": "A strong choice for per-tool-call policy and credential brokering around local coding agents, provided OS-level isolation is enough for your threat model.",
      "peers": [
        "anthropic-srt",
        "codex-cli-sandbox",
        "gemini-cli-sandbox",
        "minimal",
        "bubblewrap",
        "agents-sandbox"
      ],
      "sources": [
        "https://nono.sh",
        "https://nono.sh/docs",
        "https://nono.sh/docs/core/overview.md",
        "https://nono.sh/docs/quickstart.md",
        "https://nono.sh/docs/llms.txt",
        "https://github.com/nolabs-ai/nono"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "minimal",
      "name": "Minimal",
      "company": "Minimal",
      "hq": null,
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://minimal.dev",
      "docs": "https://minimal.dev/docs",
      "pricing_url": null,
      "github": "https://github.com/gominimal/minimal",
      "tagline": "Open-source CLI that runs coding agents in declarative, isolated sessions built from a minimal.toml and a from-source package catalogue, on Linux namespaces or a libkrun microVM.",
      "isolation": "Linux: unprivileged user and mount namespaces by default (local-minimald), or a libkrun microVM with --provider local-minvmd; macOS: always a lightweight Linux microVM booted by libkrun on Hypervisor.framework",
      "isolation_short": "namespaces or libkrun microVM",
      "isolation_class": "mixed",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": false,
      "byoc": false,
      "sdks": [
        "CLI",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Sessions can be attached, detached and kept running; the session workspace starts as a copy of the project files taken at activation; a /state directory can be shared across tasks via state_key",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Package builds get network access only when a dependency requires it and tasks \"when necessary\"; session network policy is not documented",
      "pricing": "CLI is in public beta and free to download for Linux and macOS; \"Minimal for Teams\" (managed hermetic builds, org-wide policy enforcement, attestation) is marked \"Coming Soon\"",
      "free_tier": "Free to download (public beta)",
      "compliance": null,
      "regions": null,
      "best_for": "Reproducible, declaratively scoped local sessions for Claude Code, Codex, Gemini and Pi with a supply-chain-audited toolchain",
      "overview": [
        "Minimal is \"a declarative isolation runtime for Claude Code, Gemini, Codex, Pi, and more\". You describe the environment in a `minimal.toml` kept with the code, and Minimal builds a box from a catalogue where every package is built from source with SBOM and VEX data. On Linux the default sandbox uses unprivileged user and mount namespaces (kernel 5.10+); a libkrun microVM is available with `--provider local-minvmd` and is always used on Apple Silicon macOS. Intel Macs are not supported.",
        "Sessions can be attached, detached and left running, and a hosted MCP server lets agents read the package catalogue. The CLI is in public beta (install page offers v0.7.3) and is Apache-2.0; remote execution and \"Minimal for Teams\" are marked coming soon. The README notes the full workspace does not yet build on macOS because the sandbox stack is Linux-only."
      ],
      "notable": [
        "Public beta; install page offers v0.7.3; Apache-2.0",
        "macOS uses a libkrun microVM on Hypervisor.framework; Linux defaults to unprivileged user/mount namespaces",
        "\"Remote execution is coming soon\" and Minimal for Teams is marked \"Coming Soon\""
      ],
      "pros": [
        "You want the agent environment declared in a file next to the code and rebuilt reproducibly",
        "You care about supply-chain provenance of the tools inside the sandbox",
        "You are on Apple Silicon and want a microVM rather than a container"
      ],
      "cons": [
        "Public beta; Teams and remote execution are marked coming soon; the full source workspace does not yet build on macOS",
        "Linux default sandbox uses user and mount namespaces; the libkrun microVM provider is opt-in there",
        "Intel Macs are not supported; Windows support is not documented"
      ],
      "verdict": "A public-beta CLI for developers who want declarative, reproducible agent sessions on Linux or Apple Silicon macOS.",
      "peers": [
        "agents-sandbox",
        "nono",
        "docker-sandboxes",
        "anthropic-srt",
        "microsandbox",
        "libkrun"
      ],
      "sources": [
        "https://minimal.dev",
        "https://minimal.dev/docs",
        "https://minimal.dev/docs/concepts/sandboxing",
        "https://github.com/gominimal/minimal"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "gemini-cli-sandbox",
      "name": "Gemini CLI sandboxing",
      "company": "Google",
      "hq": {
        "city": "Mountain View",
        "region": "California",
        "country": "US",
        "source": "https://policies.google.com/terms"
      },
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/google-gemini/gemini-cli",
      "docs": "https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/sandbox.md",
      "pricing_url": null,
      "github": "https://github.com/google-gemini/gemini-cli",
      "tagline": "Gemini CLI's built-in sandbox: macOS Seatbelt profiles or a Docker/Podman container around the agent's tool calls, with gVisor, LXC and a Windows option.",
      "isolation": "macOS Seatbelt (sandbox-exec) with six profiles from permissive-open (default) to strict-proxied; Docker or Podman container on any platform; gVisor (runsc), LXC/LXD (experimental) and a Windows native sandbox each on their own OS",
      "isolation_short": "Seatbelt or container",
      "isolation_class": "os",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": false,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Works on the project directory on the host; the default Seatbelt profile limits writes to the project directory",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "-open Seatbelt profiles allow direct network access, -proxied profiles route through a proxy; container sandboxes can be run with --network=none, which also blocks the Gemini API",
      "pricing": "Free, open-source CLI; the README says it is designed for Gemini Code Assist Standard/Enterprise licences with no free tier",
      "free_tier": null,
      "compliance": null,
      "regions": null,
      "best_for": "Confining Gemini CLI tool execution on a developer machine using built-in OS or container sandboxing",
      "overview": [
        "Gemini CLI ships sandboxing for the commands and tools the agent runs. It is enabled with `-s`/`--sandbox`, `GEMINI_SANDBOX` or `\"sandbox\": true` in settings. On macOS the lightweight option is the built-in Seatbelt framework via `sandbox-exec`, with `SEATBELT_PROFILE` selecting among permissive/restrictive/strict profiles in open or proxied network variants; the default `permissive-open` \"denies operations by default\" and limits writes to the project directory. Elsewhere it runs the agent in a Docker or Podman container from `ghcr.io/google/gemini-cli:latest`, replaceable with a project `.gemini/sandbox.Dockerfile`. gVisor, LXC/LXD (experimental) and a Windows native sandbox are also listed.",
        "Tool-level sandboxing is on by default (`security.toolSandboxing`). The docs note GUI applications may not work in sandboxes and that gVisor cannot reach host loopback services. The CLI is Apache-2.0; its README states it is now aimed at Gemini Code Assist Standard/Enterprise licences and that free individual access moved to Antigravity CLI on 18 June 2026."
      ],
      "notable": [
        "Six Seatbelt profiles on macOS; default permissive-open; container image ghcr.io/google/gemini-cli:latest",
        "README: free individual and Google AI Pro/Ultra access moved to Antigravity CLI on 18 June 2026; Gemini CLI targets Code Assist Standard/Enterprise",
        "Release channels: Preview (weekly), Stable (weekly), Nightly (daily)"
      ],
      "pros": [
        "You use Gemini CLI and want built-in confinement with no extra tooling on macOS",
        "You want a container-based sandbox with a custom Dockerfile on Linux",
        "You need a proxied network profile for the agent"
      ],
      "cons": [
        "Seatbelt is an OS-level sandbox on the host; container isolation is provided by Docker or Podman",
        "LXC is experimental and must be pre-created; gVisor must be selected explicitly",
        "The README says the CLI is designed for Gemini Code Assist Standard/Enterprise licences"
      ],
      "verdict": "Built-in Seatbelt or container confinement for Gemini CLI users; it is a feature of the CLI rather than a general sandbox.",
      "peers": [
        "codex-cli-sandbox",
        "anthropic-srt",
        "nono",
        "docker-sandboxes",
        "minimal",
        "gvisor"
      ],
      "sources": [
        "https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/sandbox.md",
        "https://github.com/google-gemini/gemini-cli"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "codex-cli-sandbox",
      "name": "Codex CLI sandboxing",
      "company": "OpenAI",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://openai.com/policies/terms-of-use/"
      },
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/openai/codex",
      "docs": "https://learn.chatgpt.com/docs/sandboxing",
      "pricing_url": null,
      "github": "https://github.com/openai/codex",
      "tagline": "Codex CLI's built-in sandbox: Seatbelt on macOS, bubblewrap on Linux/WSL2 and a native Windows sandbox, with read-only, workspace-write and full-access modes.",
      "isolation": "macOS: built-in Seatbelt framework; Linux and WSL2: bubblewrap (bwrap) from PATH or a bundled helper that needs unprivileged user namespaces; Windows: native sandbox with mxc process isolation (recommended), elevated lower-privilege sandbox users, or an unelevated restricted token",
      "isolation_short": "Seatbelt / bubblewrap / Windows",
      "isolation_class": "os",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": false,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Works directly on the host workspace; workspace-write confines edits to the workspace, read-only requires approval to edit or run commands",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "The sandbox controls \"whether commands can use the network\"; approval policy can \"Always ask to edit external files and use the internet\"; in ChatGPT Work, commands reach only required hostnames from a managed allowlist unless \"Allow public internet access\" is turned on",
      "pricing": "Free, open-source CLI; use requires a ChatGPT plan or API key",
      "free_tier": null,
      "compliance": null,
      "regions": null,
      "best_for": "Confining Codex CLI command execution on macOS, Linux or Windows without containers",
      "overview": [
        "Codex CLI, OpenAI's open-source coding agent \"that runs locally on your computer\", applies an OS sandbox to every command it spawns, including git, package managers and test runners. On macOS it \"works out of the box using the built-in Seatbelt framework\"; on Linux and WSL2 it needs bubblewrap (`bwrap`), using the first on PATH or a bundled helper that requires unprivileged user namespaces (Ubuntu 24.04 may need AppArmor changes); on Windows it offers a native sandbox with three implementations, `mxc` (recommended on Windows 11 24H2/25H2 builds that have it), `elevated` (dedicated low-privilege users, the \"preferred legacy fallback\") and `unelevated` (restricted token).",
        "Three modes govern scope: `read-only`, `workspace-write` (the default low-friction mode) and `danger-full-access`, which \"runs without sandbox restrictions\". Approvals (`on-request`, `never`, with an `approvals_reviewer` option) are a separate control. The Windows page does not label the sandbox experimental; the Linux mechanism documented is bubblewrap rather than Landlock/seccomp. The repository is Apache-2.0."
      ],
      "notable": [
        "Sandbox modes: read-only, workspace-write (default), danger-full-access; approval policies on-request and never",
        "Windows native sandbox with mxc (recommended), elevated and unelevated implementations; Windows 11 recommended, Windows 10 1809+ best effort",
        "Linux/WSL2 sandbox depends on bubblewrap and unprivileged user namespaces"
      ],
      "pros": [
        "You use Codex CLI and want default confinement with no containers",
        "You need a native Windows sandbox, not just WSL2",
        "You want network use gated by approvals or an allowlist"
      ],
      "cons": [
        "OS-level sandboxing on the host rather than a VM; danger-full-access removes the filesystem and network boundaries",
        "Linux requires bubblewrap and user-namespace support; Windows mxc needs specific recent builds",
        "Windows unelevated mode has weaker network isolation and no denied read paths"
      ],
      "verdict": "Built-in OS-level confinement for Codex CLI on macOS, Linux and Windows; it is a CLI feature, not a reusable sandbox.",
      "peers": [
        "gemini-cli-sandbox",
        "anthropic-srt",
        "nono",
        "minimal",
        "bubblewrap",
        "windows-sandbox"
      ],
      "sources": [
        "https://learn.chatgpt.com/docs/sandboxing",
        "https://learn.chatgpt.com/docs/sandboxing",
        "https://learn.chatgpt.com/docs/windows/windows-sandbox",
        "https://github.com/openai/codex",
        "https://github.com/openai/codex/blob/main/docs/sandbox.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "just-bash",
      "name": "just-bash",
      "company": "Vercel",
      "hq": {
        "city": "San Francisco",
        "region": "California",
        "country": "US",
        "source": "https://vercel.com/legal/modern-slavery-act"
      },
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://justbash.dev",
      "docs": "https://github.com/vercel-labs/just-bash/blob/main/packages/just-bash/README.md",
      "pricing_url": null,
      "github": "https://github.com/vercel-labs/just-bash",
      "tagline": "In-process bash reimplementation in TypeScript with a virtual filesystem, built by Vercel Labs so agents can run shell commands without touching the host.",
      "isolation": "Language-level sandbox: a TypeScript shell interpreter running inside your Node.js process with an in-memory virtual filesystem; \"The shell only has access to the provided filesystem\" and \"All execution happens without VM isolation\"",
      "isolation_short": "In-process interpreter",
      "isolation_class": "isolate",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": false,
      "sdks": [
        "JavaScript/TypeScript"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Filesystem persists across exec() calls within a Bash instance while shell state is reset per call; backends are InMemoryFs (default), OverlayFs (copy-on-write over a real directory), ReadWriteFs (direct access to a real directory) and MountableFs",
      "snapshots": null,
      "pause_resume": false,
      "gpu": false,
      "gpu_detail": "No - a JavaScript interpreter",
      "networking": "\"Network access is disabled by default\"; when enabled, limited to an allow-list of URL prefixes with GET/HEAD by default, and header transforms inject credentials at the fetch boundary",
      "pricing": "Free (Apache-2.0)",
      "free_tier": "Free",
      "compliance": null,
      "regions": null,
      "best_for": "Giving an agent a bash-like tool over a virtual filesystem inside a Node.js process, for development, testing and lightweight file manipulation",
      "overview": [
        "just-bash is \"a virtual bash environment with an in-memory filesystem, written in TypeScript and designed for AI agents\", published by Vercel Labs. It reimplements the shell and a large command set (file and text tools, `jq`, `sqlite3`, `yq`, `rg`, archives) in TypeScript, so commands never reach the host OS; optional `python3` (WebAssembly) and `js-exec` (QuickJS) are off by default. Filesystem backends range from pure in-memory to an overlay over a real directory, and `bash-tool` wraps it as an AI SDK tool. A `Sandbox` class mimics `@vercel/sandbox` for in-process development and testing.",
        "The README is explicit about the model: this is not VM isolation, Node worker limits do not reliably cap WebAssembly memory, and for arbitrary binaries it recommends Vercel Sandbox. Limits on call depth, command count, output size and time are configurable. It is labelled beta: \"This is beta software. Use at your own risk and please provide feedback.\" Requires Node 20.19+. We list it as an isolation primitive because it is a language-level sandbox, classed `isolate` here as the closest in-process category rather than a V8 isolate."
      ],
      "notable": [
        "README: \"This is beta software. Use at your own risk\"",
        "Network disabled by default; allow-list of URL prefixes when enabled",
        "README recommends Vercel Sandbox when you need a full VM or arbitrary binaries"
      ],
      "pros": [
        "You want a zero-infrastructure bash tool for agents that cannot touch the host",
        "You need deterministic, in-memory file manipulation in tests or serverless functions",
        "You want a drop-in local stand-in for Vercel Sandbox during development"
      ],
      "cons": [
        "Runs in your Node.js process; the README says \"All execution happens without VM isolation\"",
        "No arbitrary binary execution; only the reimplemented command set, plus opt-in Python and JS via WebAssembly",
        "Beta software; memory containment for WebAssembly needs external process isolation"
      ],
      "verdict": "An in-process bash interpreter with a virtual filesystem for agent tools, development and testing; its README notes that execution happens without VM isolation and points to Vercel Sandbox for a full VM.",
      "peers": [
        "vercel-sandbox",
        "langchain-sandbox",
        "pyodide",
        "webcontainers",
        "deno-sandbox",
        "wasmer"
      ],
      "sources": [
        "https://github.com/vercel-labs/just-bash",
        "https://github.com/vercel-labs/just-bash/blob/main/packages/just-bash/README.md",
        "https://github.com/vercel-labs/just-bash/blob/main/THREAT_MODEL.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "agents-sandbox",
      "name": "agents-sandbox",
      "company": "inoio",
      "hq": {
        "city": "Hamburg",
        "country": "DE",
        "source": "https://inoio.de/impressum/"
      },
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/inoio/agents-sandbox",
      "docs": "https://github.com/inoio/agents-sandbox/blob/main/docs/get-started.md",
      "pricing_url": null,
      "github": "https://github.com/inoio/agents-sandbox",
      "tagline": "GPL-3.0 launcher that gives each project and coding agent its own microsandbox microVM on Linux/KVM or Apple Silicon macOS, with secrets kept on the host.",
      "isolation": "Dedicated microsandbox microVM per project and agent with a \"separate microVM and Linux kernel\"; the README does not name the hypervisor",
      "isolation_short": "microsandbox microVM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "GPL-3.0",
      "self_host": false,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "VM is built or reused per project; agent state persists in a /home/dev volume; the project directory is mounted read-write at /workspace, or an isolated VM-internal worktree for some agents",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Egress profiles: public (default), private and host extend it, none is allowlist-only \"deny-by-default egress, not a complete air gap\"; links can open in the host browser for allowlisted schemes",
      "pricing": "Free (GPL-3.0)",
      "free_tier": "Free",
      "compliance": null,
      "regions": null,
      "best_for": "Developers who want each coding agent confined to a per-project microVM on their own Linux or Apple Silicon machine",
      "overview": [
        "agents-sandbox \"allows you to isolate your coding agents from the rest of your dev environment\" by building or reusing a microsandbox VM for each project and agent, then attaching the agent inside it. Supported agents are opencode and opencode2 (the tested defaults; opencode2 is in beta) plus pi, Claude Code, Codex, Hermes Agent and Oh My Pi, which are \"experimental - not well tested with agents-sandbox\". The host project is mounted at /workspace; other host files are invisible unless copied or mounted, and `env.secret` keeps real credential values on the host while the guest sees a placeholder.",
        "Egress is controlled by profiles from public to allowlist-only none. The README warns that the default setup copies the active agent's host configuration into the VM, which may include credentials. It runs on Linux with KVM and macOS on Apple Silicon (\"Platform requirements change; check upstream docs\") and does not need a system-wide microsandbox install. GPL-3.0."
      ],
      "notable": [
        "GPL-3.0; built on microsandbox VMs with a separate Linux kernel per project",
        "opencode2 is the default agent and in beta; five other agents are marked experimental",
        "Secrets via env.secret stay on the host; guest sees a placeholder"
      ],
      "pros": [
        "You want one microVM per project rather than one shared sandbox",
        "You need secrets kept off the guest while the agent still works",
        "You want configurable egress without a full air gap"
      ],
      "cons": [
        "Five of the seven agent profiles are marked experimental",
        "Hypervisor and boot behaviour are inherited from microsandbox and not documented here",
        "GPL-3.0; Linux/KVM or Apple Silicon only"
      ],
      "verdict": "A per-project microVM launcher for opencode users; the README marks its other agent profiles as experimental.",
      "peers": [
        "minimal",
        "microsandbox",
        "docker-sandboxes",
        "nono",
        "anthropic-srt",
        "smolvm"
      ],
      "sources": [
        "https://github.com/inoio/agents-sandbox",
        "https://github.com/inoio/agents-sandbox/blob/main/docs/get-started.md"
      ],
      "verified_on": "2026-11-14"
    },
    {
      "id": "smol-machines",
      "name": "smol machines (smolvm)",
      "company": "smol machines, inc.",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://smolmachines.com",
      "docs": "https://smolmachines.com/docs/",
      "pricing_url": "https://smolmachines.com/pricing",
      "github": "https://github.com/smol-machines/smolvm",
      "tagline": "Apache-2.0 runtime that boots OCI images as branchable libkrun microVMs on macOS, Linux and Windows, with checkpoints, portable .smolmachine files and a hosted smol cloud behind the same SDK.",
      "isolation": "Hardware-virtualized microVM with its own guest kernel for every workload: libkrun as the VMM and smol machines' libkrunfw guest kernel, on Hypervisor.framework (macOS), KVM (Linux) or the Windows Hypervisor Platform; the VMM is linked into the binary with no host daemon. smol cloud runs each machine in its own hardware-isolated microVM",
      "isolation_short": "libkrun microVM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "JavaScript/TypeScript",
        "Python",
        "Rust",
        "Go",
        "CLI",
        "REST API"
      ],
      "cold_start": "README comparison table: \"<200ms to agent ready\"; local VMs \"boot in under a second\" and a packed .smolmachine executable \"boots in under 200 ms\" (vendor claims). smol cloud machines start from warm pools",
      "cold_start_short": "<200 ms (claim)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Local machines created with `machine create` keep installed packages and state across restarts; `machine run` VMs are removed on exit. Checkpoints capture RAM, CPU state and disks and can be restored on another host or on smol cloud; branches are copy-on-write children of a live machine; a machine can be packed into a portable .smolmachine file and pushed to an OCI registry. On smol cloud the machine filesystem persists across exec and stop/start (not /tmp), and a stop exports a recovery point",
      "snapshots": true,
      "pause_resume": true,
      "gpu": true,
      "gpu_detail": "Local smolvm only: Vulkan via virtio-gpu/Venus, and CUDA API remoting over vsock to a host NVIDIA GPU (not passthrough); Vulkan is unavailable on native Windows. The docs state that hosted smol cloud does not provide GPU machines",
      "networking": "Networking is off by default; egress can be limited to named hosts, host patterns or CIDRs, and the allow list can be changed on a running machine. Credential substitution lets code use a secret bound to specific hosts while the guest sees a placeholder. Local `--auto-ports` mirrors listening guest ports onto 127.0.0.1; named networks let machines reach each other. smol cloud offers blocked, open and allowCidrs modes, published ports and optional public ingress",
      "pricing": "smol cloud: base $0.04 per running machine-hour (1-minute minimum), plus $0.05 per active vCPU-hour, $0.0162 per GB-hour of memory in use, $0.0001 per GB-hour of disk used (~$0.073/GB-month) and $0.05 per GB egress; a beta 10% discount on spend above $2,000 a month. Plans: Standard $0 (20 machines, 8 vCPU and 16 GB per machine), Pro $20 a month plus usage (100 machines, 16 vCPU, 24 GB, 200 GB disk, 99.95% SLA), Startup $150 a month plus usage (1,000 machines, 32 vCPU, 32 GB, 200 GB disk, priority support), Enterprise custom. Local and self-hosted smolvm is free",
      "free_tier": "One-time $100 sign-up credit on smol cloud, no card needed to start (up to 4 machines of up to 4 vCPU each until a card is added); smolvm is free to run locally or self-host",
      "compliance": null,
      "regions": "No region list is published; the privacy policy names Google Cloud (US regions) as compute infrastructure and says data is processed in the United States",
      "best_for": "Branching and checkpointing microVMs locally for agents and RL rollouts, then running the same machine on a hosted cloud",
      "overview": [
        "smol machines makes smolvm, an Apache-2.0 microVM runtime that boots any OCI image as a Linux VM with its own kernel, using libkrun as the VMM and its own libkrunfw guest kernel, on macOS (Apple Silicon), Linux with KVM and Windows with the Windows Hypervisor Platform. The VMM is linked into the binary, so no Docker or host daemon is needed, and memory is elastic through virtio balloon. The project describes itself as \"Branchable microVMs for AI agents\": a running machine can be branched into copy-on-write children, checkpointed with RAM and CPU state, paused and resumed, or packed into a portable .smolmachine file that runs as a self-contained executable or is pushed to an OCI registry. It is not Celesto's SmolVM (listed separately); the README says smolvm \"is not affiliated with other projects named SmolVM\".",
        "The same `Machine` API is available in TypeScript, Python, Rust and Go SDKs that embed the engine in-process or connect to smol cloud, the company's hosted service. smol cloud bills a $0.04 base per running machine-hour plus measured CPU, memory, disk and egress, with plans from $0 to $150 a month and a one-time $100 credit; idle machines can auto-stop or pause, and a paused machine with a published port resumes on its next request. Locally, smolvm adds GPU access through Vulkan (virtio-gpu/Venus) and CUDA API remoting, but the docs say smol cloud has no GPU machines. The website shows a Y Combinator (S26) badge."
      ],
      "notable": [
        "Release v1.35.4 on 18 November 2026; 7,139 GitHub stars at verification.",
        "Checkpoints capture RAM, CPU state and disks; a checkpoint taken on an arm64 Mac resumes on arm64 Linux or on smol cloud.",
        "README claims <200 ms to agent ready, and a packed .smolmachine executable boots in under 200 ms (vendor claims)."
      ],
      "pros": [
        "You want hardware-isolated microVMs that run natively on a Mac, a Linux server or Windows from the same CLI and SDK.",
        "You fan out agent tasks or RL rollouts by branching a warm, running machine instead of booting fresh ones.",
        "You want to checkpoint a running machine and move it between hosts or into a hosted cloud.",
        "You need egress allow-lists and credentials the guest code never reads."
      ],
      "cons": [
        "GPU support is local only; the docs say hosted smol cloud does not provide GPU machines.",
        "Guest architecture must match the host, and checkpoints do not restore across architectures.",
        "Compliance certifications and a hosted region list are not published.",
        "Intel macOS has no published release build and is listed as untested."
      ],
      "verdict": "An open-source libkrun microVM runtime built around branching and portable checkpoints, with a usage-billed hosted cloud on the same API.",
      "peers": [
        "smolvm",
        "microsandbox",
        "boxlite",
        "libkrun",
        "pullrun",
        "morph"
      ],
      "sources": [
        "https://smolmachines.com",
        "https://smolmachines.com/pricing",
        "https://smolmachines.com/faq",
        "https://smolmachines.com/docs/",
        "https://smolmachines.com/docs/cloud",
        "https://smolmachines.com/docs/cloud/lifecycle-storage-networking",
        "https://smolmachines.com/docs/introduction/concepts/gpu",
        "https://smolmachines.com/legal/privacy",
        "https://smolmachines.com/legal/terms",
        "https://github.com/smol-machines/smolvm",
        "https://github.com/smol-machines/smolvm/blob/main/docs/limitations.md",
        "https://github.com/smol-machines/smolvm/blob/main/docs/gpu.md",
        "https://github.com/smol-machines/smol"
      ],
      "verified_on": "2026-11-18"
    },
    {
      "id": "zeroboot",
      "name": "Zeroboot",
      "company": "Zeroboot (Adam Miribyan)",
      "hq": null,
      "kind": "open-source",
      "tier": 3,
      "alternatives": false,
      "website": "https://github.com/zerobootdev/zeroboot",
      "docs": "https://github.com/zerobootdev/zeroboot/blob/main/docs/API.md",
      "pricing_url": null,
      "github": "https://github.com/zerobootdev/zeroboot",
      "tagline": "Apache-2.0 Rust exec server that forks a pre-warmed Firecracker snapshot into a new copy-on-write KVM VM for every code run, with Python and TypeScript SDKs.",
      "isolation": "Each execution runs in its own KVM virtual machine forked from a Firecracker template snapshot: snapshot memory is mapped copy-on-write with mmap(MAP_PRIVATE) and CPU state is restored, so forks cannot see each other's writes (hardware-enforced via Intel VT-x/AMD-V)",
      "isolation_short": "CoW KVM fork of Firecracker snapshot",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "Python",
        "JavaScript/TypeScript",
        "REST API"
      ],
      "cold_start": "README (vendor benchmark): spawn latency p50 0.79 ms and p99 1.74 ms; fork plus Python exec ~8 ms; 1,000 concurrent forks in 815 ms; ~265 KB of memory per sandbox",
      "cold_start_short": "0.79 ms p50 (claim)",
      "max_session": "Each exec has a timeout_seconds parameter that defaults to 30 seconds; a maximum and long-lived sandboxes are not documented",
      "max_session_short": "30 s default timeout",
      "persistence": "None between runs: every exec forks fresh from the template snapshot. Templates (for example Python with numpy and pandas pre-loaded, or Node.js) are built once by booting Firecracker and snapshotting memory and CPU state (~15 s); template updates require a full re-snapshot",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "No networking inside forks; sandboxes communicate with the host over serial I/O only (README known limitations)",
      "pricing": "Free to self-host under Apache-2.0. A managed API at api.zeroboot.dev is offered as early access through a sign-up form; no pricing published",
      "free_tier": "Free to self-host; the README includes a demo key for the managed endpoint",
      "compliance": null,
      "regions": null,
      "best_for": "Very high-rate, short Python or Node.js snippet execution where each call gets its own VM",
      "overview": [
        "Zeroboot is an Apache-2.0 server written in Rust that runs untrusted code snippets for AI agents in VMs created by copy-on-write forking. A template is made once by booting a Firecracker VM, pre-loading a runtime and snapshotting its memory and CPU state; each request then creates a new KVM VM, maps the snapshot memory with MAP_PRIVATE and restores the CPU state, which the README benchmarks at a 0.79 ms p50 spawn (vendor benchmark). The HTTP API offers `/v1/exec` and `/v1/exec/batch` for Python (the default) and Node.js/JavaScript, plus health and Prometheus metrics endpoints, with optional API-key auth rate-limited at 100 requests per second per key. Python and TypeScript SDKs wrap the API, and it self-hosts on any Linux machine with KVM.",
        "The README calls the project a \"Working prototype\" whose fork primitive, benchmarks and API are real but \"not production-hardened yet\". Documented limitations: one vCPU per fork, no networking inside forks, userspace random-number generators that need explicit reseeding per fork because forks share the snapshot's CSPRNG state, and a full re-snapshot for template changes. Each exec has a 30-second default timeout and no state carries between runs. A managed API is in early access with no pricing published; zeroboot.dev redirects to the GitHub repository."
      ],
      "notable": [
        "Created in March 2026 by Adam Miribyan; no tagged release (the changelog lists 0.1.0 as untagged); 2,483 GitHub stars at verification.",
        "README benchmark table: 0.79 ms p50 and 1.74 ms p99 spawn, ~265 KB per sandbox (vendor benchmark).",
        "Status in the README: \"Working prototype\", \"not production-hardened yet\"."
      ],
      "pros": [
        "You run large numbers of short, independent Python or Node.js snippets and want a separate VM for each one.",
        "You want a small, self-hostable exec server with a two-endpoint HTTP API and batch execution.",
        "You want each run isolated by KVM rather than by a container or namespace."
      ],
      "cons": [
        "Labelled a working prototype that is not production-hardened.",
        "No networking inside forks and a single vCPU per fork.",
        "No state carries between runs; long-lived sandboxes are not documented.",
        "Userspace PRNGs must be reseeded per fork, and the managed API has no published pricing."
      ],
      "verdict": "A prototype snapshot-forking exec server for stateless, sub-millisecond-spawn code runs, self-hosted on Linux with KVM.",
      "peers": [
        "smol-machines",
        "firecracker",
        "e2b",
        "microsandbox",
        "riza",
        "daytona"
      ],
      "sources": [
        "https://github.com/zerobootdev/zeroboot",
        "https://github.com/zerobootdev/zeroboot/blob/main/docs/API.md",
        "https://github.com/zerobootdev/zeroboot/blob/main/docs/ARCHITECTURE.md",
        "https://github.com/zerobootdev/zeroboot/blob/main/docs/DEPLOYMENT.md",
        "https://github.com/zerobootdev/zeroboot/blob/main/CHANGELOG.md",
        "https://zeroboot.dev"
      ],
      "verified_on": "2026-11-18"
    },
    {
      "id": "matchlock",
      "name": "Matchlock",
      "company": "Matchlock (Jingkai He)",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://github.com/jingkaihe/matchlock",
      "docs": "https://github.com/jingkaihe/matchlock/tree/main/docs",
      "pricing_url": null,
      "github": "https://github.com/jingkaihe/matchlock",
      "tagline": "Experimental MIT CLI and Go, Python and TypeScript SDKs that run AI agents from OCI images in ephemeral microVMs, with a host-side proxy that enforces an egress allow-list and injects secrets the VM never sees.",
      "isolation": "MicroVM per sandbox: Firecracker on Linux with KVM and Apple Virtualization.framework on macOS (Apple Silicon); no host directories are shared with the guest, and a host-side policy and HTTP/TLS proxy sits between the guest and the network",
      "isolation_short": "Firecracker / Apple VZ microVM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": null,
      "sdks": [
        "Go",
        "Python",
        "JavaScript/TypeScript",
        "CLI"
      ],
      "cold_start": "README: a full Linux environment that \"boots in under a second\" (vendor claim); pulled images are cached for faster startup",
      "cold_start_short": "<1 s (claim)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Guest-local changes are disposable unless a persistent block volume is attached (`matchlock volume` with `run --disk @name:/guest/path`). Sandboxes started with `--rm=false` or `-d` keep running for later `exec`. There are no host-directory mounts; files move in and out with `matchlock cp`, `--copy-in`/`--copy-out` or the SDK file APIs",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Passing `--allow-host` or `--secret` seals the network so only allowed hosts are reachable, and the allow list can be edited at runtime; real secrets stay on the host and are injected into allowed outbound requests while the guest sees a placeholder. SDK interception rules can mutate requests and responses. The Go SDK blocks private IP ranges by default when a network config is sent; `--no-network` gives a guest with no NIC; ports can be published or forwarded. Linux uses a transparent proxy (nftables DNAT on ports 80/443); macOS uses Virtualization.framework NAT, or gVisor's userspace TCP/IP stack when intercepting",
      "pricing": "Free and open source under MIT; no hosted offering is documented",
      "free_tier": "Free (MIT)",
      "compliance": null,
      "regions": null,
      "best_for": "Running coding agents on a Linux server or Apple Silicon Mac in disposable microVMs with locked-down egress and API keys kept on the host",
      "overview": [
        "Matchlock is an MIT-licensed CLI, with Go, Python and TypeScript SDKs, for \"running AI agents in ephemeral microVMs\". It boots OCI images (pulled from a registry, imported from `docker save` or built from a Dockerfile with BuildKit inside a VM) as microVMs: Firecracker on Linux with KVM and Apple Virtualization.framework on Apple Silicon Macs, according to the repository's developer reference. The README says a sandbox \"boots in under a second\" and that the CLI behaves the same on a Linux server and a MacBook. The examples directory covers Claude Code, OpenAI Codex CLI, Docker inside the sandbox and browser automation with Playwright MCP.",
        "Its focus is the network boundary. With an allow-list or a secret configured, a host-side proxy blocks every host not on the list and substitutes real credentials into outbound requests in flight, so code in the VM only ever reads a placeholder; SDK hooks can also rewrite requests and responses. Host directories are never shared: files are copied in and out explicitly, and only attached block volumes persist. The README marks the project \"Experimental\" and subject to breaking changes; it was at v0.3.1 at verification, and no hosted service is documented."
      ],
      "notable": [
        "Release v0.3.1 on 31 October 2026; 654 GitHub stars at verification.",
        "Secrets are injected by a host-side MITM proxy, so the guest environment only holds a placeholder value.",
        "`matchlock build` runs BuildKit inside a VM and imports the result into the local image store."
      ],
      "pros": [
        "You want an agent to have a full Linux VM while its API keys never enter the guest.",
        "You need a default-deny egress allow-list you can edit while the sandbox runs.",
        "You run agents on both Linux servers and Apple Silicon Macs and want one CLI and SDK for both.",
        "You want no host-directory sharing, with explicit copy-in and copy-out."
      ],
      "cons": [
        "Labelled experimental and subject to breaking changes.",
        "No host-directory mounts; files must be copied in and out or kept on block volumes.",
        "Snapshots, pause/resume, GPU support and session limits are not documented.",
        "Self-host only; no hosted service is documented."
      ],
      "verdict": "An experimental local microVM sandbox for agents whose distinguishing feature is host-side egress control and secret injection.",
      "peers": [
        "smol-machines",
        "microsandbox",
        "nono",
        "pullrun",
        "boxlite",
        "docker-sandboxes"
      ],
      "sources": [
        "https://github.com/jingkaihe/matchlock",
        "https://github.com/jingkaihe/matchlock/blob/main/AGENTS.md",
        "https://github.com/jingkaihe/matchlock/tree/main/docs",
        "https://github.com/jingkaihe/matchlock/blob/main/docs/lifecycle.md",
        "https://github.com/jingkaihe/matchlock/releases"
      ],
      "verified_on": "2026-11-18"
    },
    {
      "id": "sandboxd",
      "name": "sandboxd",
      "company": "sandboxd",
      "hq": null,
      "kind": "open-source",
      "tier": 2,
      "alternatives": true,
      "website": "https://sandboxd.io",
      "docs": "https://sandboxd.io/what-is-sandboxd",
      "pricing_url": "https://sandboxd.io/managed",
      "github": "https://github.com/tastyeffectco/sandboxd",
      "tagline": "MIT self-hosted engine that runs each app or agent in its own hardened Docker container on one server, with built-in coding agents, preview URLs, idle stop and wake, and per-task git checkpoints.",
      "isolation": "One hardened runc Docker container per app (--cap-drop=ALL, no-new-privileges, read-only root filesystem with tmpfs /tmp, memory, PID and file-descriptor limits, non-root user); gVisor (runsc) can be switched on for every sandbox with SANDBOXD_RUNTIME=runsc",
      "isolation_short": "Hardened runc container (gVisor opt-in)",
      "isolation_class": "container",
      "open_source": true,
      "license": "MIT",
      "self_host": true,
      "byoc": true,
      "sdks": [
        "REST API"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Each sandbox's workspace is a bind-mounted directory on the host that survives stop and reboot; the container's writable layer is read-only and /tmp is tmpfs. Idle sandboxes are stopped with docker stop after 35 minutes by default and started again on the next request, so files persist but memory and running processes are not kept. Snapshot, fork and restore endpoints exist but are labelled experimental on directory storage",
      "snapshots": true,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Each sandbox's dev server gets a preview URL (s-<id>-<port>.preview.<domain>) routed by Traefik; previews are public and plain HTTP by default, with TLS as a config switch. Egress is open by default in the self-hosted build: sandboxes sit on a Docker bridge and can reach the internet, the LAN and cloud metadata endpoints; an nftables egress subsystem is compiled off and a per-domain allowlist is on the roadmap. Model credentials are injected by a control-plane proxy so they never enter the sandbox",
      "pricing": "Free to self-host under MIT. Managed (sandboxd installs and runs it on your own server): $199 one-time setup and $79/month ($49/month for the first 10 founding customers). sandboxd Cloud (a public HTTPS URL for a server you run): $9.99 per server per month, $4.99 with the 50% launch discount. Compute is always your own server",
      "free_tier": "Free to self-host",
      "compliance": null,
      "regions": null,
      "best_for": "Running many agent-built apps or agent workspaces on one server you control, each in its own container with a live preview URL",
      "overview": [
        "sandboxd describes itself as \"the open-source engine behind AI app builders\". One Go control plane drives Docker on a single host, Traefik routes preview URLs and SQLite holds state, with a reconciler that converges Docker to the database on boot. Each app gets one sandbox: a hardened runc container with its own workspace, an in-sandbox supervisor (runtimed) and a live preview URL. Coding agents run inside the sandbox (the README lists OpenCode, Claude Code, Codex and Pi; OpenCode works keyless on a free tier), and a credential-injecting proxy keeps model keys out of the container. Before every agent task the workspace is checkpointed to a private git ref, so a task can be reverted. Everything is exposed through an authenticated /v1 REST API with SSE task streams, file read and write, exec and git endpoints; the web console is an optional API client.",
        "The Security page states that \"sandboxd runs AI-generated code you did not write\" and that the default shared-kernel container is \"a strong isolation boundary and a weak security boundary\". gVisor is available as an opt-in runtime (the vendor puts syscall-heavy work at about 1.7–4× slower), and Firecracker or Cloud Hypervisor microVMs are on the roadmap. Egress is open by default. Idle sandboxes are stopped after 35 minutes and woken on the next request, which keeps files but not memory. The project is pre-1.0 (v0.3.x) and single-server; multi-host scheduling is listed under \"Later\". Self-hosting is free; the vendor sells a Managed service on your own server (BYOC) and a Cloud tunnel that gives your server a public URL, and does not host compute."
      ],
      "notable": [
        "v0.3.26 released 31 October 2026; 997 GitHub stars at verification; the README labels it \"Beta · 0.x\".",
        "Configuration docs: /v1 auth is on by default (console session or bearer API key, no loopback bypass) and the API binds to 127.0.0.1:9090.",
        "Agent tasks are checkpointed to refs/sandboxd/checkpoints/<task> before they run and can be reverted with one API call."
      ],
      "pros": [
        "You want a self-hosted, MIT-licensed control plane for many sandboxes with preview URLs on one ordinary server.",
        "You need coding agents inside each sandbox without putting model keys in the container.",
        "You want per-task checkpoints and one-call revert of an agent's changes.",
        "You want idle sandboxes to stop and wake on the next request so one host holds many apps."
      ],
      "cons": [
        "The default boundary is a shared-kernel container; the vendor recommends gVisor for less-trusted code and one host per trust domain for mutually distrusting tenants.",
        "Egress is open by default; a per-sandbox domain allowlist is still on the roadmap.",
        "Single Docker host only; memory is not kept when an idle sandbox stops, and snapshots are experimental on directory storage.",
        "Pre-1.0: minor versions may carry breaking changes."
      ],
      "verdict": "A self-hosted engine for app-builder and agent platforms that runs each workload in a hardened container on one server; its own docs call the default container a weak security boundary and point to gVisor or separate hosts for untrusted tenants.",
      "peers": [
        "opensandbox",
        "microsandbox",
        "daytona",
        "coder",
        "openhands-runtime",
        "gvisor"
      ],
      "sources": [
        "https://sandboxd.io",
        "https://sandboxd.io/what-is-sandboxd",
        "https://sandboxd.io/concepts",
        "https://sandboxd.io/security",
        "https://sandboxd.io/reference/architecture",
        "https://sandboxd.io/reference/configuration",
        "https://sandboxd.io/reference/api/",
        "https://sandboxd.io/guides/agents",
        "https://sandboxd.io/guides/tasks",
        "https://sandboxd.io/guides/runtime-and-store",
        "https://sandboxd.io/roadmap",
        "https://sandboxd.io/managed",
        "https://sandboxd.io/cloud",
        "https://sandboxd.io/privacy",
        "https://github.com/tastyeffectco/sandboxd",
        "https://github.com/tastyeffectco/sandboxd/releases"
      ],
      "verified_on": "2026-11-18"
    },
    {
      "id": "gondolin",
      "name": "Gondolin",
      "company": "Earendil Inc.",
      "hq": null,
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://earendil-works.github.io/gondolin/",
      "docs": "https://earendil-works.github.io/gondolin/quickstart/",
      "pricing_url": null,
      "github": "https://github.com/earendil-works/gondolin",
      "tagline": "Apache-2.0 local Linux micro-VMs for agent-generated code, driven from a TypeScript SDK or CLI, with network egress, secrets and file mounts controlled by host-side JavaScript policy.",
      "isolation": "Local Linux micro-VM per sandbox: libkrun where available (macOS on Apple silicon, Linux x86_64 with read/write /dev/kvm) and QEMU otherwise (Intel Macs, Linux arm64, Linux x86_64 without KVM via TCG); networking and the virtual filesystem run in a host daemon (gondolind) outside the guest",
      "isolation_short": "libkrun / QEMU micro-VM",
      "isolation_class": "microvm",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": false,
      "byoc": false,
      "sdks": [
        "JavaScript/TypeScript",
        "CLI"
      ],
      "cold_start": "Docs: VMs \"boot in under a second\" on a Mac or Linux machine (vendor claim); no measurement methodology given",
      "cold_start_short": "<1 s boot (claim)",
      "max_session": null,
      "max_session_short": null,
      "persistence": "Each VM starts from a clean copy of its image and the root disk is discarded on close by default (root disk modes readonly, memory or cow); disk-only checkpoints capture the writable root disk as a .qcow2 file that many VMs can resume from in parallel; tmpfs paths and VFS mounts are not in checkpoints, and durable data is meant to go to a host-backed mount such as /workspace",
      "snapshots": true,
      "pause_resume": false,
      "gpu": null,
      "gpu_detail": null,
      "networking": "No NAT: the host daemon terminates guest traffic. HTTP/1.x and HTTPS are parsed and performed by the host (TLS intercepted with a local CA the guest trusts), with any public host allowed by default and loopback, private, link-local and metadata ranges blocked; other TCP and non-DNS UDP are dropped. allowedHosts / --allow-host restrict egress, and JavaScript hooks or daemon-side policies can filter and rewrite requests; secrets reach the guest only as placeholders that the host replaces for allowed hosts. HTTP/2, HTTP/3 and QUIC are not supported",
      "pricing": "Free (Apache-2.0)",
      "free_tier": "Free",
      "compliance": null,
      "regions": null,
      "best_for": "Running agent-generated code in disposable local micro-VMs where egress, credentials and file access must stay under programmable host-side policy",
      "overview": [
        "Gondolin, built by Earendil, runs code from AI agents inside \"lightweight micro-VMs\" on a Mac or Linux machine: libkrun on Apple silicon and on Linux x86_64 with KVM, and QEMU on other hosts. You create a VM from the TypeScript SDK (`VM.create()`, `vm.exec()`) or open a shell with `npx @earendil-works/gondolin bash`. The network stack and virtual filesystem run on the host in a small daemon and are programmable from JavaScript: HTTP(S) allowlists and request/response hooks, placeholder-based secret injection so the guest never sees the real token, and custom VFS mounts. An ingress gateway exposes guest HTTP services, SSH is supported, and the daemon can run on a remote machine while policy and credentials stay local.",
        "The docs target a run-persist-discard lifecycle per agent turn or task. Root disks are thrown away by default; disk checkpoints let you prepare an environment once and resume many VMs from it, but the limitations page states that Gondolin \"does not provide full VM save/restore\" of RAM and process state. Guest images are Alpine-based (~200 MB+, downloaded on first use), and the image builder supports only Alpine. Requires Node.js 22.19 or newer. The docs call it \"an early project\"; the latest tagged release on GitHub is v0.14.0 (October 2026)."
      ],
      "notable": [
        "Docs: \"this is an early project\"; latest GitHub release v0.14.0 (22 October 2026)",
        "Secrets are injected by the host only for allowed destinations; the guest sees a placeholder",
        "Disk-only checkpoints (.qcow2) can be resumed many times in parallel; no RAM or process save/restore"
      ],
      "pros": [
        "You want a microVM boundary around agent code on your own Mac or Linux machine without a cloud service",
        "You need egress allowlists and request hooks written in JavaScript rather than fixed profiles",
        "You want credentials usable by the guest without ever placing them inside it",
        "You want to prepare a base environment once and start many disposable VMs from a checkpoint"
      ],
      "cons": [
        "Docs describe it as an early project",
        "No full VM save/restore; checkpoints capture the root disk only",
        "Only Alpine guest images; HTTP/2, HTTP/3, QUIC and generic UDP are not supported through the network mediation",
        "TypeScript SDK only; requires Node.js 22.19+ (a static CLI is also available)"
      ],
      "verdict": "A local micro-VM sandbox whose distinctive feature is host-side, JavaScript-programmable control of network, secrets and files; the docs call it an early project.",
      "peers": [
        "matchlock",
        "microsandbox",
        "smolvm",
        "boxlite",
        "agents-sandbox",
        "libkrun"
      ],
      "sources": [
        "https://earendil-works.github.io/gondolin/",
        "https://earendil-works.github.io/gondolin/quickstart/",
        "https://earendil-works.github.io/gondolin/workloads/",
        "https://earendil-works.github.io/gondolin/snapshots/",
        "https://earendil-works.github.io/gondolin/network-access/",
        "https://earendil-works.github.io/gondolin/limitations/",
        "https://github.com/earendil-works/gondolin",
        "https://earendil.com"
      ],
      "verified_on": "2026-11-18"
    },
    {
      "id": "agent-safehouse",
      "name": "Agent Safehouse",
      "company": "Agent Safehouse (eugene1g)",
      "hq": null,
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://agent-safehouse.dev",
      "docs": "https://agent-safehouse.dev/docs/",
      "pricing_url": null,
      "github": "https://github.com/eugene1g/agent-safehouse",
      "tagline": "Apache-2.0 Bash script that runs Claude Code, Codex and other coding agents under a deny-first macOS sandbox-exec policy, granting the project directory and denying SSH keys, credentials and other repos.",
      "isolation": "macOS Seatbelt via sandbox-exec: composable policy profiles assembled into a deny-first policy around the agent process on the host kernel; the docs describe it as \"a host-level macOS policy wrapper, not a VM boundary\"",
      "isolation_short": "sandbox-exec (Seatbelt)",
      "isolation_class": "os",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": false,
      "byoc": false,
      "sdks": [
        "CLI"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Works directly on the host filesystem: the selected workdir (the invocation directory by default, or --workdir) is read/write, extra paths can be granted read/write or read-only, and installed toolchains are readable",
      "snapshots": null,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Network access is open by default \"for registries, APIs, remotes, and MCP servers\"; the docs list data exfiltration over the network from readable files among what Safehouse does not fully protect against",
      "pricing": "Free (Apache-2.0)",
      "free_tier": "Free",
      "compliance": null,
      "regions": null,
      "best_for": "Confining local coding agents on a Mac to the project directory with near-zero overhead and no VM or container",
      "overview": [
        "Agent Safehouse \"sandboxes LLM coding agents on macOS so they can only access paths they need\". It is a single self-contained shell script (also installable with Homebrew) that assembles composable `sandbox-exec` profiles into a deny-first policy and runs the agent inside it, for example `safehouse claude --dangerously-skip-permissions`. The site lists Claude Code, Codex, OpenCode, Amp, Copilot CLI, Gemini CLI, Aider, Goose, Cursor Agent, Cline and others, with per-agent investigation pages in the docs.",
        "By default the selected invocation directory is read/write (Safehouse \"does NOT auto-widen the default workdir to an enclosing Git repo\"), system and toolchain paths are readable, and SSH private keys, shell startup files, browser profile data, the clipboard and cloud credential stores are denied unless opted in with `--enable` flags. Environment variables are sanitized and network access is open. The project calls itself \"a hardening layer, not a perfect security boundary\" and recommends running it inside a VM when stronger isolation is needed. macOS only; the latest GitHub release is v0.13.1 (November 2026)."
      ],
      "notable": [
        "Single self-contained Bash script; no dependencies beyond Bash and macOS",
        "Deny-first: SSH keys, other repos, browser data and cloud credentials denied unless explicitly enabled",
        "README: \"a hardening layer, not a perfect security boundary against a determined attacker\""
      ],
      "pros": [
        "You run coding agents in permissive or yolo modes on a Mac and want the blast radius limited to the project",
        "You want native toolchains to keep working with no VM, container or workspace sync",
        "You want per-feature opt-ins (Docker, Keychain, clipboard, Xcode) instead of all-or-nothing access"
      ],
      "cons": [
        "macOS only",
        "OS-level policy on the host kernel; the docs state it is not a VM or hypervisor boundary",
        "Network is open by default and egress filtering is not documented"
      ],
      "verdict": "A zero-dependency, deny-first sandbox-exec wrapper for local coding agents on macOS; it reduces blast radius rather than providing VM-grade isolation.",
      "peers": [
        "container-use",
        "nono",
        "anthropic-srt",
        "codex-cli-sandbox",
        "gemini-cli-sandbox",
        "minimal"
      ],
      "sources": [
        "https://agent-safehouse.dev",
        "https://agent-safehouse.dev/docs/",
        "https://agent-safehouse.dev/docs/overview.html",
        "https://agent-safehouse.dev/docs/default-assumptions.html",
        "https://agent-safehouse.dev/docs/options.html",
        "https://agent-safehouse.dev/docs/isolation-models.html",
        "https://github.com/eugene1g/agent-safehouse",
        "https://github.com/eugene1g/agent-safehouse/releases"
      ],
      "verified_on": "2026-11-18"
    },
    {
      "id": "container-use",
      "name": "Container Use",
      "company": "Dagger, Inc.",
      "hq": null,
      "kind": "local",
      "tier": 2,
      "alternatives": true,
      "website": "https://container-use.com",
      "docs": "https://container-use.com/quickstart",
      "pricing_url": null,
      "github": "https://github.com/dagger/container-use",
      "tagline": "Experimental Apache-2.0 MCP server and CLI from Dagger that gives each coding agent a fresh container on its own git branch, for review with log, diff and terminal before you merge, apply or delete.",
      "isolation": "A container per environment, built and run through Dagger (Docker required), paired with a dedicated git branch via Git worktrees; each command runs \"inside a NEW container within the environment\"",
      "isolation_short": "Dagger container",
      "isolation_class": "container",
      "open_source": true,
      "license": "Apache-2.0",
      "self_host": false,
      "byoc": false,
      "sdks": [
        "CLI",
        "MCP server"
      ],
      "cold_start": null,
      "cold_start_short": null,
      "max_session": null,
      "max_session_short": null,
      "persistence": "Each environment's changes and command history are tracked on its own git branch and survive across chats; an agent can resume an existing environment by its ID with the container state reused; an environment_checkpoint MCP tool saves an environment's current state as a container image to a registry destination",
      "snapshots": true,
      "pause_resume": null,
      "gpu": null,
      "gpu_detail": null,
      "networking": "Background commands and services can expose ports, returning an environment-internal and a host-external address; egress policy is not documented",
      "pricing": "Free (Apache-2.0)",
      "free_tier": "Free",
      "compliance": null,
      "regions": null,
      "best_for": "Running several coding agents in parallel on one repository, each in its own container and branch, and reviewing their work with git before merging",
      "overview": [
        "Container Use, from Dagger, \"gives your coding agents sandboxed dev environments to work in, so they can run safely in parallel\". It is an open-source MCP server that also works as a CLI (`container-use` or `cu`): you add `container-use stdio` as an MCP server to Claude Code, Cursor, Goose, Codex, Gemini CLI or another MCP-compatible agent, and each agent session creates a fresh environment from your current branch. An environment combines a container, built with Dagger on Docker from a configurable base image (Ubuntu 24.04 by default) plus setup and install commands, with a dedicated git branch that records every change and command.",
        "You review work with `container-use log`, `diff`, `terminal` or `checkout`, then `merge` it, `apply` it as staged changes, or `delete` the environment. Secrets are configured as references (1Password, environment variables, HashiCorp Vault or files), resolved inside the container and stripped from logs and outputs. The README carries an \"experimental\" stability badge and says the project \"is in early development\"; the latest tagged release is v0.4.2 (August 2025), while the repository continues to receive commits."
      ],
      "notable": [
        "README badge: stability experimental; \"This project is in early development and actively evolving\"",
        "Latest tagged release v0.4.2 (19 August 2025); repository still active",
        "Secret references are resolved inside the container and stripped from logs and command output"
      ],
      "pros": [
        "You want to run multiple coding agents on one repo without them overwriting each other's work",
        "You want to review agent work with ordinary git branches, diffs and merges",
        "You want to drop into an agent's live container to debug when it gets stuck",
        "You use any MCP-compatible agent rather than one specific vendor"
      ],
      "cons": [
        "Labelled experimental; no tagged release since v0.4.2 in August 2025",
        "Requires Docker and Git on the host; isolation is a container rather than a VM",
        "Network egress controls are not documented"
      ],
      "verdict": "An experimental, git-centred way to give each local coding agent its own container and branch; best suited to parallel agent work you review before merging.",
      "peers": [
        "docker-sandboxes",
        "agents-sandbox",
        "minimal",
        "nono",
        "openhands-runtime",
        "agent-safehouse"
      ],
      "sources": [
        "https://container-use.com",
        "https://container-use.com/quickstart",
        "https://container-use.com/environment-workflow",
        "https://container-use.com/environment-configuration",
        "https://container-use.com/secrets",
        "https://container-use.com/agent-integrations",
        "https://github.com/dagger/container-use",
        "https://github.com/dagger/container-use/blob/main/mcpserver/tools.go",
        "https://github.com/dagger/container-use/releases",
        "https://dagger.io/privacy-policy"
      ],
      "verified_on": "2026-11-18"
    }
  ]
}