Giving an AI agent real access to your computer usually means one of two things: approving every single step until you stop reading, or setting up heavy isolation most people skip.
For the Meta Global AI Developer Hackathon I built barebox, a small sandbox you can start for an agent in under a second. The idea is simple. Every action gets one of three letters, like file permissions:
→ r: reading and looking things up. It just happens.
→ w: changes that can be undone. Kept on an undo layer you keep or throw away at the end.
→ x: things that can't be taken back, like paying, sending or deleting. It waits for your yes.
And you pick one of three modes, like any command-line tool:
→ ask: it checks with you before any change
→ auto: only the irreversible things need you (the default)
→ trust: it runs on its own
Payments ask in every mode.
It's built for spinning up agents on the fly: no setup, no root, no Docker image, about 8 MB each, up to 40 side by side on one laptop.
In the demo, Meta's muse model books a trip inside barebox. It searched, compared and planned on its own, and stopped only for the 8 things that couldn't be undone: payments, bookings, a calendar entry, messages and an email. The agent never saw the real payment key.
Honest limit: it's a guardrail against mistakes and tricked agents, not a fortress. Open source under Apache-2.0.
https://github.com/hamr0/ba...