Google Workspace for your AI agents.
Give Claude Code, Codex, Cursor, and other MCP-capable coding agents one shared place to coordinate work, retain decisions, and create clear handoffs.
Concord MCP · Quick start · Website · Contribute
Open several coding agents in one checkout and their chat sessions cannot tell each other who owns a file, which assumptions guide the work, or whether a task is ready for review. Concord adds that shared work layer to the repository.
Agents register their presence, claim files and modules before editing, attach decisions to a task, and hand work off with tests and review evidence. Humans inspect the same state through the CLI and dashboard.
| Concord records | Your team can inspect |
|---|---|
| Agent presence | Who is active, away, blocked, or waiting for review |
| Work claims | Which files and modules each task expects to touch |
| Task memory | Decisions, assumptions, findings, questions, and blockers |
| Handoffs | Changes, tests, risks, provenance, and open review questions |
concord-mcp is our open-source,
local-first MCP server. It gives coding agents a shared SQLite work-state inside
the repository they are changing.
npm install -g @concord-ai/concord-mcp
concord installconcord install adds the agent instructions for the current repository. Connect
the MCP server using the guide for
Claude Code,
Codex, or
Cursor.
Concord is early and under active development.
| Tool | Role |
|---|---|
register_agent |
Add an agent instance to the live roster |
get_work_state |
Read agents, active tasks, overlaps, and review state |
claim_work |
Record expected files and modules before edits begin |
update_task |
Attach progress, decisions, questions, and findings |
get_task_context |
Resume a task from its ordered shared context |
handoff |
Capture changes, tests, risks, and review evidence |
The CLI exposes the same state to developers:
concord status
concord dashboard
concord tasks
concord handoff <task-id>Concord keeps its source of truth in .concord/concord.db at the repository
root. It also generates readable artifacts for handoff and review:
.concord/
├── concord.db
├── HANDOFF.md
├── REVIEW_PACKET.md
└── WORK_STATE.json
The server stays model-agnostic and works with MCP clients that can run a local process. Your code and task content stay in the checkout.
Concord MCP uses the MIT license. The codebase enables TypeScript strict mode, keeps modules focused, and requires each pull request to stay under 600 lines.
- Browse good first issues
- Read the contribution guide
- Open or follow issues
- Explore the source
getconcord.ai · npm · GitHub
