Agent Safehouse

Agent Safehouse is a macOS native sandbox for local AI coding agents. It wraps an agent process in kernel level enforcement through the operating system's sandbox facility, so an agent running with its permission prompts switched off still cannot write outside the project directory or read sensitive files. A destructive command aimed at the home folder fails with an operation not permitted error before any file is touched. It is built for developers who let agents run autonomously on their own Mac and want a hard boundary rather than a prompt based one. The access model is deny first: by default the selected working directory, the git root, is read and write, installed toolchains are readable, and most of the home directory, including SSH keys, cloud credentials and other repositories, is denied. Extra directories can be granted read only or read and write, and shell functions make every agent launch inside the sandbox automatically. The project states it has been tested with Claude Code, Codex, OpenCode, Amp, Copilot CLI, Gemini CLI, Aider, Goose, Auggie, Pi, Cursor Agent, Cline, Kilo Code and Droid. Agent Safehouse is free and open source under the Apache 2.0 licence, with no paid plan or pricing page. It installs through Homebrew or as a single self contained shell script, with no build step and no dependencies beyond Bash and macOS. The tool is macOS only, since it relies on the platform's own sandboxing rather than containers or virtual machines.