eugene1g/agent-safehouse

Health v2 · maintenance only

Adoption not part of the score

2037 stars · 88 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Agent Safehouse is a macOS-native sandboxing tool that runs local LLM coding agents (Claude Code, Codex, Gemini CLI, Aider, etc.) inside a deny-first sandbox built on sandbox-exec/Seatbelt. It ships as a single Bash script or Homebrew formula with composable policy profiles that grant agents only the file and integration access they need.

Use cases

• run claude code with --dangerously-skip-permissions without risking my home directory
• sandbox an AI coding agent so it can't read ~/.ssh or ~/.aws
• restrict an LLM agent to read/write only the current project directory
• prevent an AI agent from running rm -rf on my files
• create a custom least-privilege sandbox profile for my own agent
• safely try out new coding agents on my macbook
• kernel-enforced file access limits for autonomous agents

When to choose

• you run AI coding agents on macOS and want kernel-enforced protection of files outside your project
• you want to use agents in permissive/yolo modes with reduced blast radius
• you need composable, auditable sandbox profiles without extra dependencies beyond Bash and macOS

When to avoid

• you need sandboxing on Linux or Windows (macOS-only via sandbox-exec)
• you require a hard security boundary against a determined malicious attacker - it is a hardening layer, not a perfect isolation
• your workflow depends on agents accessing many paths outside the workdir and you don't want to manage policy rules

Facets

cli-tool · maturity active

security cli developer-tools security developer-tools cli sandboxing seatbelt sandbox-exec llm-agents least-privilege claude-code coding-agents shell-script deny-first sandbox ai-agents command-line macos

Member repositories

For agents

markdown · JSON · MCP: product_card(name="eugene1g/agent-safehouse")