Concord MCP Server: Let Claude Code, Codex and Cursor Agents Work Together

• Claims a task and warns you if another agent is already on it

  Marks the agent as present and claims the files or area it is about to change. If another agent already holds part of that area, the answer says so before any edits happen. It also accepts work that was assigned to this agent.

  Try: “Before you edit src/app/page.tsx, claim it and tell me if anyone else is working there.”

  Write · start_work

• Shows who is working, on what, and what messages are waiting

  Reads the shared state: which agents are present, which tasks and claims exist, and which of them look stale. It can also open one agent's inbox or outbox, or one prompt and reply thread. It only reads and is marked read-only.

  Try: “Who else is working in this repo right now, and are any claims stale?”

  Read · inspect_work

• Saves notes on a task or sends a live message to another agent

  Adds decisions, assumptions or findings to a task so they outlast the session. It can also send a prompt to another agent, such as one in Codex or Cursor, or reply to one. Delivery fails with an error if that agent cannot be reached.

  Try: “Tell the Codex agent I will keep the page and it should build the API under src/app/api.”

  Write · update_work

• Assigns, hands off, accepts or releases a task between agents

  Changes who owns a task. Assigned work stays unowned until the named agent accepts, and a handoff offer leaves ownership with the sender until the other side accepts. Every change is logged and checked against the task's version number, so two agents cannot both win.

  Try: “Hand the login fix to the reviewer agent with the test results attached.”

  Write · transfer_work

• Records the result and marks a task ready for review or done

  Stores what was done and the evidence for it, then either leaves the task open as a handoff, marks it ready for review or closes it. The saved record feeds the handoff and review files in the repo's .concord folder.

  Try: “Mark the checkout refactor review-ready and list the tests I ran.”

  Write · finish_work

Watch out for

• Setup edits agent config and adds hooks

  concord setup writes MCP entries, instruction files and hooks for several agents, including a Claude Code PreToolUse check that runs before tool calls. It merges into existing files and is safe to re-run, but review the changes in git or your home folder. github.com

• The server can repair agent plugins and hooks when it starts

  On start, the server runs concord adapters ensure in the background for the agent app it was launched from. It changes files only if you opted in to automatic adapters during concord setup (otherwise it just prints a notice). When on, it can write: a link to the bundled plugin (Claude Code, a link in ~/.gemini/extensions/concord-relay for Gemini CLI, ~/.cursor/extensions/get-concord-ai.concord-relay plus ~/.cursor/hooks.json for Cursor); Codex MCP entry and hooks in config.toml, then it starts and enables the Codex app-server daemon with remote control; a Grok Build plugin via grok plugin install --trust; a Goose config.yaml entry; Copilot CLI ~/.copilot/mcp-config.json and hooks; Devin CLI MCP config and user hooks; Antigravity ~/.gemini/config/mcp_config.json and hooks. concord adapters status shows what is installed. github.com

• Agents can send each other prompts

  update_work delivers text straight into another agent's session, so a message from one agent becomes instructions for another. Only run Concord with agents you trust. github.com

• Writes a .concord folder in your repo

  Tasks, a handoff file, review packet and event log are saved in .concord at the repo root. Setup adds a .gitignore there so it is not committed by default. github.com

• Sends anonymous usage statistics

  Telemetry goes to getconcord.ai. Set CONCORD_TELEMETRY_DISABLED=1 or DO_NOT_TRACK=1 to switch it off. github.com

Security advisories