Bothread
A calm, local room for your AI Agents to work together
One thing that's annoyed me every time I've tried running more than one coding agent on the same project: they have no idea the other exists. Two agents touch the same file, one silently overwrites the other, and you find out when the diff looks wrong.
Bothread is a free, open-source, local app that puts multiple AI coding agents into one shared "room" over MCP. Before an agent edits anything, it claims the files it needs (exclusive or shared, glob-based, with a TTL). If another agent tries to claim something overlapping, the claim is denied and both agents see it happen live — no silent collisions. They also share a live chat thread, a task board, and can hand off files to each other. A human (you) watches the room and can pause, mute, revoke, or approve any agent's action at any time. Everything's local: SQLite for state, an append-only audit log, no API keys required to run the coordination layer itself.
It's not tied to any one vendor — it works with Claude Code, Cursor, Antigravity, Gemini, Codex, and OpenCode. In the demo I ran, one of the three agents was driven through OpenCode using a free model, alongside Claude Code and Antigravity (Gemini), and all three split up the work, claimed files, coordinated in chat, and built a working platformer together — a real "collision PREVENTED" fired live mid-build, and one agent caught and fixed a bug that made the game unwinnable while the others kept working.
To try it: npx bothread start (zero-install; no API keys needed to run the room itself — you still need whatever credentials your chosen agents normally require), or npm install -g bothread once and then run bothread start from anywhere.
Genuine question for this sub: for those of you running local models as agents, is file-level claiming (whole-file locks) granular enough, or would you actually want symbol/function-level locking given how local models tend to make smaller, more frequent edits?
1 Comment
The coordination layer is the interesting innovation here, not multi-agent coding itself. I'd keep validating whether teams are really paying to avoid file conflicts or to make parallel development trustworthy enough that they can confidently use multiple agents on the same codebase. That's a much bigger problem than locking files.