Agent Safehouse

Sandbox your LLM coding agents on macOS so they can only access the files and integrations they actually need.

Agent Safehouse uses sandbox-exec with composable policy profiles and a deny-first model. It includes profiles for major coding agents and app-hosted agent workflows while keeping normal development usage practical.

Install

Homebrew:

brew install eugene1g/safehouse/agent-safehouse

Standalone script:

mkdir -p ~/.local/bin
curl -fsSL https://github.com/eugene1g/agent-safehouse/releases/latest/download/safehouse.sh \
  -o ~/.local/bin/safehouse
chmod +x ~/.local/bin/safehouse

Philosophy

Agent Safehouse is designed around practical least privilege:

• Start from deny-all.
• Allow only what the agent needs to do useful work.
• Keep developer workflows productive.
• Make risk reduction easy by default.

It is a hardening layer, not a perfect security boundary against a determined attacker.

HOME access by default

HOME_DIR is used to render precise home-relative rules in the assembled policy. By itself, it does not grant recursive read access to your home directory.

WORK_DIR similarly exposes the selected workdir to built-in and appended policy rules through workdir-literal, workdir-subpath, and workdir-prefix. Relative helper arguments begin with /, letting a reusable policy target files such as <workdir>/.env without hardcoding an absolute project path.

Default Safehouse behavior is narrower:

• metadata-only traversal on /, the path to $HOME, and $HOME itself so runtimes can probe explicitly allowed home-scoped paths
• directory-root reads for ~/.config and ~/.cache so tools can discover XDG locations
• a few explicit home-scoped files/directories from always-on profiles, such as git/ssh metadata and shared agent instruction folders

In practice, stat "$HOME" can succeed while ls "$HOME" and cat ~/secret.txt still fail unless a more specific rule grants that path.

If you want to remove even the default home exceptions, use --append-profile; appended profiles load last, so their deny rules can narrow earlier defaults.

Built-In System Path Resolution

Safehouse's built-in profiles/* modules may include macOS compatibility paths such as /etc, /private/etc/resolv.conf, or /private/etc/localtime.

At policy render time, Safehouse resolves built-in absolute paths from allow file-read* rules and emits matching grants for the real target path when the authored path is a symlink. That keeps host-specific system files working without broadening the source profiles to recursive /private/etc access.

Current scope is intentionally limited to built-in absolute literal and subpath read grants. User-provided path grants still normalize separately, and writable or metadata-only built-in rules are not auto-expanded by this mechanism today.

Documentation

• Website: agent-safehouse.dev
• Docs: agent-safehouse.dev/docs
• Policy Builder: agent-safehouse.dev/policy-builder

Machine-Specific Defaults

If you keep shared repos, caches, or team folders in machine-specific locations, keep those settings out of project config and put them in a shell wrapper plus a local appended profile.

This lets you define your own sane defaults once and reuse them from claude, codex, amp, or app launchers:

POSIX shells (zsh / bash):

# ~/.zshrc or ~/.bashrc
export SAFEHOUSE_APPEND_PROFILE="$HOME/.config/agent-safehouse/local-overrides.sb"

safe() {
  safehouse \
    --add-dirs-ro="$HOME/server" \
    --append-profile="$SAFEHOUSE_APPEND_PROFILE" \
    "$@"
}

safe-claude() { safe claude --dangerously-skip-permissions "$@" }

fish:

# ~/.config/fish/config.fish
set -gx SAFEHOUSE_APPEND_PROFILE "$HOME/.config/agent-safehouse/local-overrides.sb"

function safe
    safehouse \
      --add-dirs-ro="$HOME/server" \
      --append-profile="$SAFEHOUSE_APPEND_PROFILE" \
      $argv
end

function safe-claude
    safe claude --dangerously-skip-permissions $argv
end

Example machine-local policy file:

;; ~/.config/agent-safehouse/local-overrides.sb
;; Host-specific exceptions that should not live in shared repo config.
(allow file-read*
  (home-literal "/.gitignore_global")
  (home-subpath "/Library/Application Support/CleanShot/media")
  (subpath "/Volumes/Shared/Engineering")
)

;; Keep root environment files unavailable even though the workdir is writable.
(deny file-read* file-write* (workdir-literal "/.env"))

Use --add-dirs-ro or --add-dirs for normal shared-folder access, and keep --append-profile for machine-local policy exceptions or final deny/allow overrides. That pattern is useful when the repo is shared but each developer machine has different local mount points.

Git worktrees are auto-detected at launch when the selected workdir itself is a Git worktree root: that worktree gets the shared Git metadata access it needs when its common dir lives outside the selected workdir, and the other existing linked worktrees for that repo become readable by default for cross-tree inspection. That snapshot does not update for already-running processes, so if you create worktrees under a stable parent such as ~/worktrees, prefer adding that root explicitly with --add-dirs-ro.

All detailed documentation (setup, usage, options, architecture, testing, debugging, and investigations) lives in the VitePress docs site.

Linux Alternatives

Safehouse is specifically tailored for macOS (sandbox-exec). For developers looking for Linux-native sandboxing for AI coding agents, consider:

• vetto — Zero-daemon kernel sandbox (Landlock LSM & seccomp-bpf on Linux, Seatbelt on macOS) with automated PATH shims (vetto enable <agent>).
• bubblewrap — Unprivileged user namespaces sandboxing utility for Linux.
• firejail — Mature SUID sandbox with ready-made profiles for common applications.
• nono.sh — Landlock plus seccomp-notify sandboxing with privilege elevation without restart.
• sandlock — Pure-Python sandbox combining Landlock, seccomp-bpf and seccomp user notification; no root, containers or C compiler needed.
• isolated-agent — Minimalist helper for isolating a single agent run.