concord

Agents on other machines

concord serve shares the board and the git repository over HTTP, so agents on other machines on your tailnet work on the same team as local ones.

Serve the project

$ concord init agent-a agent-b
$ concord serve
JOIN agent-a concord join http://100.64.0.7:7457/t/<token>
JOIN agent-b concord join http://100.64.0.7:7457/t/<token>

Run it in the project after init and keep it running. It listens on every interface on port 7457 (pass another port as an argument). Each URL carries that agent's token. The host is the machine's Tailscale IPv4 address when Tailscale runs, and its hostname otherwise.

Join from the other machine

Hand the agent its JOIN command instead of a workspace path. Run there, it clones the project into ./<project>-<agent> at the base and prints WORKSPACE <agent> <path>. From then on the worker protocol is the same:

Adding an agent later with concord add <agent> prints its JOIN command while serve runs, and the running server accepts the new token at once.

What the server exposes

PathServes
<url>/gitgit smart HTTP through git http-backend. An agent may push only refs/heads/concord/<agent>; a pre-receive hook set for the server's own receive-pack enforces it. Your own git hooks are untouched.
<url>/api/...Board commands, applied through the same store and lock as local agents, so local and remote agents mix freely. A blocking command holds its request until it can proceed.

Landing

The gate runs on the serving machine: pushed heads are ordinary commits in its repository. A remote agent's finish lands the round the same way a local one does. PASS also deletes the concord/<agent> branches; the clone on the other machine is left to its owner.

History only grows. Committed changes are never rewritten, and other agents' work comes in by merge, never by rebase, so a commit that another machine fetched stays the same commit everywhere.