Concord MCP
Live messaging and shared work-state for coding agents — detect overlapping work before edits, share decisions, and hand off tasks with evidence across harnesses.
- Source repo
- Get-Concord-AI/concord-mcp
- Stars
- ★ 1.4k
- Last updated
- 1d ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 75/100 · Good
At a glance
- Works with
- Universal · cross-platformCodex · Claude Code
- You'll need
- Typical use
- A developer running Claude Code and Codex in parallel on the same repo who wants agents to detect file-claim collisions before editing
- Main limitation
- No silent rerouting: delivery fails immediately when a named agent has no reachable endpoint, so you must monitor adapters status
- Website
- getconcord.ai ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v0.23.0 (2026-11-17)+44.1 ★/day over 7 days
- Source review
- 75/100 · Good
What does this agent do, and when should you use it?
Concord MCP is an open-source, local-first communication and coordination layer for AI coding agents. Through a single MCP server, it lets agents such as Claude Code, Codex, Cursor, Gemini CLI, and Grok Build discover each other, exchange replyable direct prompts, claim files and tasks, and retain task memory after sessions end. All shared state lives in a local SQLite database under .concord/ at the repository root, alongside human-readable artifacts like HANDOFF.md and REVIEW_PACKET.md. It is explicitly not an orchestrator, code reviewer, or autonomous agent — it is the shared work-state layer around the agents you already run. Agents use five MCP tools, while humans and CLI-oriented agents access the same workspace through the concord command line.
After installation, concord setup creates a local .concord/ workspace, registers the MCP server for Claude, Cursor, Gemini, Grok, and Codex, and writes tool instructions into client configs. Agents collaborate through five MCP tools: start_work registers presence, claims a task, and reports scope overlaps before editing; inspect_work reads workspace/task state, an inbox/outbox, and durable prompt/reply threads; update_work records context or sends live messages via operation: "prompt"/"reply" with to_agent_id and an idempotency_key; transfer_work handles assignment, acceptance, handoffs, and reassignment; finish_work records evidence and marks tasks review-ready, complete, or closed. Writes carry an agent_id to keep presence live, lifecycle changes use expected_version for optimistic concurrency, and every ownership change is retained in an append-only audit history. The CLI offers concord status, who, tasks, handoff, review-packet, and a live dashboard TUI.
- A developer running Claude Code and Codex in parallel on the same repo who wants agents to detect file-claim collisions before editing
- Teams that want task decisions and context to survive session ends instead of disappearing into chat history
- Handing off a task from one agent to another with full handoff evidence and explicit acknowledgment
- Independent reviewers who need a REVIEW_PACKET.md with scope, tests, risks, and provenance before approving work
- Solo developers monitoring agent presence, tasks, alerts, and activity in real time via concord dashboard
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
npm install -g @concord-ai/concord-mcpnpm install -g @concord-ai/concord-mcp@latest
How do you install or deploy this agent?
Install globally via npm and initialize inside your repository:
bash
npm install -g @concord-ai/concord-mcp
cd /path/to/your/repository
concord setupNo API keys are required. setup creates the local .concord/ workspace, registers MCP servers, and merges instructions into existing configs — safe to re-run. Restart your agent clients once afterward to enable live delivery. Flags --no-adapters, --no-mcp, and --require-adapters adjust behavior.
How do you use this agent?
After restarting client sessions, ask two agents to work in the same repository; they will discover each other and exchange messages. Humans can inspect at any time:
bash
concord status
concord dashboard
concord tasks
concord handoff <task-id>
concord review-packet <id>
concord adapters statusUse --repo or --workspace (mutually exclusive globals) to select a repository from anywhere. There is no universal /concord slash command; clients work through MCP tools plus installed instructions. To upgrade:
bash
npm install -g @concord-ai/concord-mcp@latest
concord --versionWhat are this agent's strengths and limitations?
- Direct cross-harness messaging between Claude Code, Codex, Cursor, Gemini CLI, and Grok Build without a human relaying context
- Local SQLite source of truth with .concord/ gitignored by default — code and task content never leave the machine
- Optimistic concurrency via expected_version plus append-only audit history prevents conflicting state transitions
- Strong operational tooling: live TUI dashboard, adapters status reachability matrix, and stale-claim detection
- No silent rerouting: delivery fails immediately when a named agent has no reachable endpoint, so you must monitor adapters status
- Only effective within a shared local checkout — it is not a hosted sync service or multi-repo solution
- Sends telemetry to getconcord.ai by default (disablable); server-side IP and country-code storage has no automatic expiry
- Delivery requires installing adapters and restarting client sessions, and live delivery depends on each harness's state
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Concord MCP This agent | 75 · Good | — | ★ 1.4k | 1d ago | TypeScript | Codex · Claude Code |
| Vibe Ecosystem Coding Runtime | 76 · Good | CLIFree | ★ 553 | 9d ago | C# | Claude Code |
| MCO | 75 · Some gaps | CLIFree + model costs | ★ 558 | 14d ago | Python | Codex · Claude Code |
| AWS Agent Toolkit | 61 · Some gaps | Agent plugin / skillFree + model costs | ★ 3.5k | today | Python | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The code and README describe a local SQLite default and an optional CONCORD_ALLOWED_ROOTS path restriction. However, ordinary startup opens a workspace and registers identity automatically, so least privilege is only partly established (2). Setup modifies several client configuration and instruction files. It is described as merging and rerunnable, but the packet does not show per-change confirmation, so confirmation is partial (2). The README and SECURITY.md specify default telemetry fields, receiver-side IP/country code, and opt-out controls; src/index.ts also shows distinct local and cloud paths, supporting strong data-flow transparency (3). The docs say code and task content are not sent, but receiver-side IP/country data has no automatic expiry, leaving a sensitive-data handling gap (2). Dependencies are listed, but the packet provides no dependency audit, lockfile contents, or vulnerability mitigation evidence (1). Default telemetry, update checks, client config writes, and optional cloud communication are disclosed, though external effects and retention leave residual concerns (2). Cloud unlink and disabling global mode offer reversible paths, and old data is not automatically deleted; cleanup scope for all setup writes is not fully shown in this sample (2). Package metadata, license, and a security reporting channel provide attribution, but publisher identity is not registry-verified and maintenance ownership is not very clear (1).
The README’s tools, versioned state transitions, and delivery limits broadly align with the entry point and artifact-writing code; the multiple cloud, global, and local modes create boundaries that still need verification (2). Node requirements, package manager, install paths, and a cross-platform CI matrix are documented, but the source sample cannot establish availability of every runtime dependency (2). Startup errors are written to stderr with a nonzero exit code, while some client-root resolution errors silently fall back to the working directory, so failure feedback is not consistently complete (2).
The product targets many coding agents, MCP clients, and CLI users, with corresponding entry points and scenarios described in the README (3). It clearly describes itself as a shared work-state and coordination layer rather than an autonomous agent, and documents hook-only pull limitations and immediate failure when no endpoint is reachable (3). Supported clients and task behavior are concrete, but some behavior depends on client sessions, identity, and integration setup, so trigger conditions require user understanding (2). Local, global, cloud, Git worktree, multiple OS, and Node-version cases are addressed; the sample does not enumerate every compatibility constraint across these combinations (2).
The README’s quick start, features, CLI, privacy section, linked guides, and architecture description provide clear information structure (3). npm and standalone installation, Node requirements, changed config files, upgrade steps, and major options are documented (3). The five workflow tool names are stable and correspond across docs and source entry points; migration compatibility cannot be fully determined from this sample (2). Demo, CLI examples, and client guide links are present, but the sample has no FAQ answers or broad troubleshooting section (2). Non-goals, delivery limits, telemetry contents, and pre-1.0 status are explicit (3). The MIT text and package metadata agree (3). Package version, upgrade command, and release workflow are documented, but no changelog content is included for review (2). CONTRIBUTING, an engineering guide, security reporting instructions, and release workflow exist; the responsible maintenance party and identity remain insufficiently established (2).
Structured workflow tools, audit events, handoffs, and review artifacts are directly usable by later agents and humans; sample artifact tests support the intended formats (3). Cross-harness messaging, collision warnings, and durable task state address a clearly described coordination gap and offer substantial marginal value (3). Installation changes several client configs, telemetry is enabled by default, and some features depend on adapters or cloud; local operation and opt-out controls reduce the burden, but user setup and maintenance costs remain (2).
The README gives concrete telemetry fields, configuration names, and workspace-root behavior that can be checked against the entry-point code (3). README, SECURITY.md, package.json, CI workflow, and sample tests corroborate some license, version, telemetry, and artifact behavior claims; many product claims remain documentation-only in the packet (2). The docs distinguish local behavior, client limits, and product intent and disclose receiver-side data, but privacy properties such as irreversible pseudonyms lack implementation detail in the supplied material (2).
- Telemetry is sent to getconcord.ai by default. It excludes code and task content, but the receiver stores request IP and country code without automatic expiry.
- concord setup writes or updates several client configuration and instruction files; review those changes and their cleanup scope before using it in shared or managed environments.
- Publisher identity is not verified by the FollowAgents registry, and this static packet contains no dependency vulnerability audit evidence.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What's Changed * feat: add Hermes MCP config writer with isolated tests by @bconnnnn in https://github.com/Get-Concord-AI/concord-mcp/pull/258 * feat: search past work with inspect_work query by @alexechoi in https://github.com/Get-Concord-AI/concord-mcp/pull/375 * feat: replies expect no answer unless asked (expects_reply) by @alexechoi in https://github.com/Get-Concord-AI/concord-mcp/pull/376 * chore: prepare release 0.23.0 by @alexechoi in https://github.com/Get-Concord-AI/concord-mcp/pull/377 Full Changelog: https://github.com/Get-Concord-AI/concord-mcp/compare/v0.22.0...v0.23.0View on GitHub ↗ -
Release notes excerpt
What's Changed * feat: identify OpenCode sessions from MCP request metadata by @bferanmi806-sketch in https://github.com/Get-Concord-AI/concord-mcp/pull/308 * feat: opencode support by @Seym0n in https://github.com/Get-Concord-AI/concord-mcp/pull/306 * fix: detect OpenCode 2 versions and stop status checks from starting its service by @alexechoi in https://github.com/Get-Concord-AI/concord-mcp/pull/343 * docs: list OpenCode in the translated READMEs by @alexechoi in https://github.com/Get-Concord-AI/concord-mcp/pull/344 …View on GitHub ↗ -
Release notes excerpt
What's Changed * docs: add translated READMEs by @alexechoi in https://github.com/Get-Concord-AI/concord-mcp/pull/337 * fix: separate agent heartbeat, activity and reported status by @akramcodez in https://github.com/Get-Concord-AI/concord-mcp/pull/224 * fix: launch managed Codex through resolved Windows shims by @dailyun in https://github.com/Get-Concord-AI/concord-mcp/pull/330 * fix: pin the agent kind in Goose, Grok, Gemini and Cursor MCP entries by @akramcodez in https://github.com/Get-Concord-AI/concord-mcp/pull/317 …View on GitHub ↗
FAQ
Is Concord MCP open source?
Is Concord MCP safe to use?
Is Concord MCP still maintained?
Who makes Concord MCP?
What is Concord MCP's official website?
Thanks — we use this to improve the page.