Concord MCP

Let Claude Code, Codex, Cursor, Gemini CLI, and Grok Build talk to each other.

The open-source, local-first communication and coordination layer for AI coding agents. Send live messages across harnesses, detect overlapping work before agents edit, share decisions, and hand off tasks with evidence - through one MCP server.

Demo · Quick start · Supported agents · Website · Contributing

See it work

Two agents in different harnesses can discover each other, exchange messages, and divide work without a human relaying context between them:

Claude Code → Concord   Claim src/app/page.tsx
Codex       → Concord   Claim src/app/page.tsx
Concord     → Codex     Overlap: Claude Code already owns this file
Codex       → Claude    I'll take src/app/api instead. Does that work?
Claude      → Codex     Yes. I'll keep the page and use your API contract.

Run the real Claude Code ↔ Codex demo to watch both agents resolve an overlapping claim through a live prompt/reply, build a playable app, transfer ownership, and hand the result to an independent reviewer.

Quick start

npm install -g @concord-ai/concord-mcp
cd /path/to/your/repository
concord setup

Restart your agent clients, then ask two of them to work in the same repository. Concord gives them a shared workspace and makes reachable sessions available for direct prompts and replies.

Supported agents

There is no universal /concord slash command — commands are client-specific. Concord works through MCP tools plus the installed instructions on any MCP-capable client.

Live delivery depends on the receiving harness and session state. Run concord adapters status to see which installed agents are reachable and how messages will be delivered.

Communication is the starting point

Messaging gets agents talking. Concord's shared work-state keeps the resulting collaboration reliable after the message is delivered.

Concord is not another autonomous agent or orchestrator. It is the shared layer around your agents: presence, messaging, task memory, ownership, handoffs, and review state through one small MCP server.

The tools

Writes accept an agent_id, which keeps presence live just by working. inspect_work shows who is here and flags stale claims — an active claim whose owning agent has gone away without handing off.

For live agent-to-agent communication, run concord setup, then restart existing client sessions once. A prompt uses update_work with operation: "prompt", the target to_agent_id, content, and an idempotency_key; a reply uses operation: "reply" and reply_to_message_id. A receipt-bearing adapter steers a busy turn or starts an idle turn. Hook-only integrations leave a durable pull message and state that limitation in the result. Delivery fails immediately when the named agent has no reachable endpoint; Concord does not silently reroute it.

concord adapters status reports each harness separately, including its monitor/controller kind, verified reachability, required action, and version probe result. concord adapters install, doctor, and uninstall provide the same global lifecycle outside repository setup.

Concord resolves the repository workspace automatically. Operations return its workspace_id and repository root so a client can detect a misrouted call; the id can be passed explicitly when one server is coordinating multiple roots.

Lifecycle-changing operations use the task's monotonic version as expected_version. If two agents act on the same version, only the first transition succeeds. Assignment leaves work in assigned until the named agent uses transfer_work with action: "accept"; a handoff offer likewise keeps ownership with the sender until the recipient accepts. Every ownership change is retained in an append-only audit history.

What you get

SQLite is the local source of truth, kept in the .concord/ at the root of the repo the work is happening in. The MCP server resolves that root from CONCORD_REPO_ROOT if set, then Claude Code's CLAUDE_PROJECT_DIR (which Claude Code sets automatically, even for a user-scoped server), then its working directory — so every agent in one repo shares one store. Set CONCORD_REPO_ROOT when running the server somewhere its working directory is not inside the repo.

Linked Git worktrees follow Git's commondir metadata to the primary checkout, so the main checkout and all linked worktrees intentionally share one Concord database and workspace id.

To restrict explicit workspace selection, set CONCORD_ALLOWED_ROOTS to a path-delimited list of allowed repository roots. Without an allowlist, decoded roots must still exist and be directories.

concord setup adds .concord/ to the repository's .gitignore, so the generated workspace stays local by default. Teams that want selected artifacts in PRs can remove that rule or force-add the human-readable files:

.concord/
├── concord.db          local source of truth
├── HANDOFF.md          human-readable handoff
├── REVIEW_PACKET.md    review-ready evidence
└── WORK_STATE.json     generated export (optional)

CLI

Concord supports both typed MCP tools and a regular CLI. MCP-capable agents can call the tools directly; humans and CLI-oriented agents can work with the same shared workspace through concord commands.

Read the full README on GitHub →