Run a project manager agent and a backend engineer agent in separate terminals, coordinating tasks through a shared board.
Use structured debate mode to have two AI agents formally argue a design decision and get a judge verdict saved to memory.
Build a shared second-brain knowledge graph that multiple AI agents add to and search across sessions.
Assign finished agents to idle status and reassign them new tasks without restarting the server.
Requires two Python dependencies: the MCP protocol library and a file-locking library.
Claude Team MCP is a Python server that lets multiple AI coding agents work together as a team on the same project. Instead of running one AI assistant at a time in a single terminal, you can run several at once, each with its own role, such as a project manager, a backend engineer, or a QA reviewer, and have them communicate through a shared channel. The coordination happens through a single shared state file that all connected agents can read and write to. Any MCP-compatible client tool can join the same team by pointing at the same file: the README mentions Claude Code, Cursor, Codex CLI, and Gemini CLI. The shared state includes a task board, a chat channel, and a persistent memory store that survives across sessions. One of the more distinctive features is a structured debate mode. When agents disagree on a design decision, one can start a formal debate: participants submit proposals, critique each other with explicit agree or disagree reasoning, revise their positions only when presented with a convincing argument, and a designated judge issues a final verdict. The result is saved to project memory. There is also a second-brain system, a small knowledge graph stored on disk with notes, links, tags, and a daily journal. Agents can add to it and search it independently of the main team channel. On Windows, agents can open new terminal windows for additional agents and close them programmatically when work is done. The server is written in Python and requires two dependencies: the MCP protocol library and a file-locking library. File writes are done atomically with a temp-file-then-rename approach, so multiple agents writing at the same moment do not corrupt the shared state. The project also tracks agent availability, so finished agents flip to idle and can be reassigned new tasks.
← shalinda-j on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.