eugene1g/agent-safehouse

What it solves

Agent Safehouse provides a security layer for LLM coding agents running on macOS. It prevents agents from accessing files and integrations they don't need, reducing the risk of accidental or malicious file system access by restricting agents to a "deny-first" model of least privilege.

How it works

The tool uses macOS's sandbox-exec to create a sandbox environment. It employs composable policy profiles that start with a total deny-all approach and explicitly allow only the necessary paths and actions. It includes built-in profiles for major coding agents and handles macOS-specific system path resolution (such as resolving symlinks for system files) to ensure the tool remains functional without granting overly broad access.

Who it’s for

Developers using AI coding agents on macOS who want to harden their environment and ensure their agents only have access to the specific project directories and home directory metadata required for their work.

Highlights

• Deny-first security model: Starts from a total block and only allows specific, required access.
• Composable policies: Supports custom profiles and appended profiles for machine-specific overrides.
• ** macOS specific**: Tailored specifically for sandbox-exec on macOS.
• Git worktree detection: Automatically detects and grants access to shared Git metadata when using Git worktrees.
• Git/SSH metadata access: Includes always-on profiles for essential metadata to keep agents functional.