Put work in. Get shipped code out.
The hands-off end of the spectrum. Drop a task in the hopper and walk away — CodeyBox burns tokens until the work is polished, production-grade, and tested, then audits and merges it. Finished work hops out the other side.
Queue it. Walk away.
A task is just a title and a prompt against one of your repos. CodeyBox picks the right agent, runs it in a throwaway VM, and moves it through an atomic Plan → Work → Audit → Merge pipeline. Changes land cleanly or not at all. Work items run in parallel, spend is tracked per agent, and durable state survives restarts — so it's safe to leave running while you're out at lunch.
Autonomous, and safe to leave alone.
Polished to production-grade
Agents don't just make it compile — CodeyBox keeps working until the change is tested and polished enough to ship, not just a first draft.
Multi-agent fleet
Routes each task to whichever agent is best and available — Claude Code, Codex, Copilot, Cursor, Gemini, opencode and more — with quota-aware fallback when one hits a limit.
Isolated microVMs
Each phase runs in a fresh KVM-backed microVM — not a container — behind a host-enforced egress firewall with least-privilege, scoped credentials.
Audit gate before merge
Nothing lands until it passes secret scanning, SAST, and an LLM security review. Clean merges happen host-side; agents are only invoked for genuine conflicts.
Cost tracking built in
See exactly where your subscription spend goes, per agent, per task — so unattended runs never turn into a surprise.
Yours to extend
Add agents, auditors, forges, or credential backends behind clean interfaces — no forking. Config-driven auditors mean managed repos can be any stack.
One queue, the whole team.
CodeyBox isn't a private black box. Everyone on the team can watch the task queue, follow each run through Plan → Work → Audit → Merge, and read the outputs and spend as they happen. Anyone can drop work in; everyone can see exactly where it stands.
Wired into the rest of the suite.
- Monitor CodeyBox jobs from inside Agnes, alongside your hands-on sessions. planned
- Accept tasks delegated from Agnes — and hand the finished, merged work straight back. planned
- Take work straight from a JobTrack board and report progress back to it as the queue moves. planned
Fill the hopper.
CodeyBox is open source and MIT licensed. Point it at a repo, queue some work, and go grab lunch.
View CodeyBox on GitHub