control-room

control-room is a way for one person to run many AI coding agents on one repository at once. It is a protocol, a set of prompts and a few small scripts, under Apache-2.0. The agents can be Claude Code, Codex or any agent that works in a git checkout: in the cloud, in an e2b sandbox, or locally in worktrees.

It came out of daily use: one person, a local coordinator and six cloud workers building one product. Parallel agents fail in dull ways. A decision waits in a chat nobody has open. A push lands on the wrong branch. So each role gets a short list of what it never does.

• The owner, a person, answers the desk, starts the workers and holds the credentials.
• The coordinator, a local agent session, splits goals into epics, writes briefs, and is the only writer of the tracker and the only one who lands on the integration branch.
• A worker takes one epic from a brief, pushes every turn on its own branch, and ends with a pull request and one line: READY or BLOCKED.

The desk is the point of the design: one place where every decision that needs the person waits, either a static page or GitHub Issues.

Today you get the protocol, English and Russian prompts, a guard that refuses a worker's pushes to other branches, force pushes and merges, a handoff schema and its validator, a script that posts local checks as commit statuses so they cost no Actions minutes, runners, a Claude plugin, and an epic replayed end to end. The lessons page records the failure behind each rule.

Not built yet

It is early, at v0.3.4. The protocol and prompts are in daily use. The scripts and runners are still being generalised from the project they came from. The e2b runner's live path has not been run with a real key. The Codex mapping is written from Codex's documentation, and marks what it could not verify. The end-to-end example is a replay on a toy repository; a real worker's transcript comes next.