Agent Safehouse corrals coding agents inside sandbox

This is the right kind of AI tooling: not a smarter agent, but a tighter blast radius for the ones developers already trust with real code.

• –It targets the practical failure mode of local agents: accidental reads, writes, and tool access outside the repo.
• –The deny-first approach is stronger than after-the-fact cleanup because it blocks overreach before it happens.
• –`sandbox-exec` makes the idea immediately useful on macOS, but also keeps it clearly platform-specific.
• –The project looks most valuable as a wrapper around existing agents, since it improves safety without forcing teams to change editors or models.
• –The repo’s own framing matters: it’s a hardening layer, not a perfect security boundary, which is the honest pitch.