agent-safehouse

Agent Safehouse is designed to address the security risks posed by AI coding agents (such as Claude, Cursor, etc.) when running locally. It leverages macOS's sandbox-exec tool to create a strict sandbox environment for these AI agents, which defaults to denying all file access and only permits access to the specific files and integrations necessary for the agent to complete its tasks. This prevents AI from accidentally or maliciously reading or modifying sensitive files.

Usage is highly flexible. You can install it with a single command via Homebrew or download it via script. At runtime, simply prepend the safehouse command and a corresponding configuration profile (e.g., safehouse claude) to the command used to launch your AI agent, placing it into the sandbox. It also supports dynamically adding or restricting access paths via command-line arguments and allows users to create local configuration files to override or append to the default rules.

This tool is ideal for developers, operations engineers, and agent developers who need to run AI coding agents securely on their local machines. Whether for daily development, reviewing AI-generated code, or running automated scripts, it can be used to isolate risks. For team collaboration or managing multiple project environments, it also helps establish a unified security baseline.

The recommended approach is to treat it as an important hardening layer, not an absolute security boundary. It's advised to start with the default configuration and gradually add allow rules based on the AI agent's actual needs (e.g., accessing specific project directories, Git repositories). For highly sensitive environments, you can use the --append-profile parameter to further tighten the default rules, such as removing exceptions like metadata access to the $HOME directory.