A sandbox and crew harness for AI coding agents. Your code stays on the host; the agent stays in its room. Most of the other work on this page was built inside it.
- An actual boundary. One container per repository, non-root, with an iptables and ipset allowlist that rejects everything but the handful of hosts the agent legitimately needs. A
.sandboxignoremasks the files it should never see — secrets read as empty, directories as bare. - Crews, not a single assistant. Several agents, each in its own container, addressing each other by mention in rooms, with floor control so N agents do not talk forever. A teammate's message is untrusted data; only the human's is authoritative.
- An access ladder. Write to its own subtree, read-only elsewhere, or caged — network sealed to its own model endpoint, no credentials, no memory. The human commits; agents never do.
- Model is not harness. A member declares a model and gets the native tool for it — Claude Code, Codex, or OpenCode — so the roster is not locked to one vendor.
- The unglamorous half. Named resumable sessions, a clipboard bridge so images paste through the container wall, desktop notifications, 1Password resolution across multiple accounts, and a dashboard to watch and steer the whole thing.
MIT licensed. This website was built, deployed and debugged from inside it.