Agent Safehouse
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. A zero-dependency, deny-first sandbox-exec wrapper for local coding agents on macOS; it reduces blast radius rather than providing VM-grade isolation.
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"
Agent Safehouse specifications
| Attribute | Details |
|---|---|
| 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" |
| Deployment | Local / OS-level |
| Open source | Yes (Apache-2.0) |
| SDKs and interfaces | CLI |
| Cold start (vendor claim) | Not documented |
| Max session length | Not documented |
| 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 |
| GPU | Not documented |
| 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 | Not documented |
| Regions | Not documented |
| Company headquarters | Not documented |
Facts from Agent Safehouse’s public documentation, verified November 18, 2026. Performance numbers are vendor claims.
Strengths and trade-offs
Strengths
- 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
Trade-offs
- 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
Good fit for: Confining local coding agents on a Mac to the project directory with near-zero overhead and no VM or container.
How Agent Safehouse compares
| Sandbox | Company HQ | Isolation | Open source | Self-host / BYOC | Cold start (vendor) | Max session | Snapshots / pause | GPU | SDKs |
|---|---|---|---|---|---|---|---|---|---|
| Agent Safehouse | Not documented | sandbox-exec (Seatbelt) | Yes | No / No | Not documented | Not documented | Not documented / Not documented | Not documented | CLI |
| Container Use | Not documented | Dagger container | Yes | No / No | Not documented | Not documented | Yes / Not documented | Not documented | CLI, MCP server |
| nono | Not documented | Landlock / Seatbelt | Yes | No / No | Not documented | Not documented | Yes / Not documented | Not documented | CLI, Rust, Python, JavaScript/TypeScript, Go |
| Anthropic Sandbox Runtime | United States | OS-level (Seatbelt / bubblewrap) | Yes | Yes / Not documented | Process start | Not documented | No / No | Not documented | CLI (srt), TypeScript (npm) |
| Codex CLI sandboxing | United States | Seatbelt / bubblewrap / Windows | Yes | No / No | Not documented | Not documented | Not documented / Not documented | Not documented | CLI |
| Gemini CLI sandboxing | United States | Seatbelt or container | Yes | No / No | Not documented | Not documented | Not documented / Not documented | Not documented | CLI |
| Minimal | Not documented | namespaces or libkrun microVM | Yes | No / No | Not documented | Not documented | Not documented / Not documented | Not documented | CLI, MCP server |
Head-to-head: Agent Safehouse vs Container Use
Frequently asked questions
What is Agent Safehouse?
How does Agent Safehouse isolate code?
Is Agent Safehouse open source?
How much does Agent Safehouse cost?
What are the alternatives to Agent Safehouse?
Sources
Check before you rely on it. Products change quickly, and this page reflects public documentation as of November 18, 2026. Confirm anything that matters in the provider’s own documentation.
Cite this page
Agent Sandboxes. (2026). Agent Safehouse. agentsandboxes.org. Last reviewed November 18, 2026. https://agentsandboxes.org/sandboxes/agent-safehouse/