Run AI coding agents in a sandbox with a read-only filesystem.
curl -fsSL safe-claude.com/install.sh | bashcd your-project
safe-claude
safe-codex
safe-opencodeOnly the current directory and supported agent config files are writable. Everything else is read-only.
safe-opencode sets OPENCODE_PERMISSION to allow all permissions.
- Linux or macOS
- Linux: firejail
- macOS: sandbox-exec (built-in)
- Node.js + npm
MIT