crew-mcp

Run Claude Code, Codex, Antigravity (Gemini), and local models as one coding crew from the conversation you already use.

• Use the CLI logins and subscriptions you already have.
• Send the same diff to independent models and compare their judgment.
• Keep write runs isolated and decide what lands on your branch.

Crew is an MCP server plus captain playbooks. Your host CLI becomes the captain; other agents become workers and reviewers.

Install in 60 seconds

Requires Node.js 20+, git, and an authenticated Claude Code or Codex CLI.

npm install -g crew-mcp
crew-mcp install --target all
crew-mcp verify

Restart your host CLI, then ask naturally:

Have Codex implement the rate limiter, then have Claude review it.

crew-mcp verify checks the installed skill and tool-catalog parity. See the installation guide for source installs, project scope, Antigravity setup, Codex wake modes, and uninstalling.

Choose a workflow

| You ask | Crew runs | Use it for | | --- | --- | --- | | "Have Codex build this." | Dispatch — one agent, one worktree, one result | A bounded implementation, investigation, or review | | "Have three models review this." | Panel — parallel reviewers, full review each | Independent judgment on the same target | | "Keep iterating until it ships." | Iterate — criteria-gated build/review rounds | Work that has to meet a standard, not just compile |

Runs are asynchronous. The captain reports the dispatch immediately, and a watcher wakes the conversation when supported; you can keep using the host CLI while workers run.

Iterate to acceptance

crew-iterate turns a request into a criteria-gated delivery loop:

The confirmed criteria become the contract for the implementer and every reviewer. Crew combines dispatched reviewers with a host-native review, scores each criterion PASS/FAIL, and folds findings back through continue_run. The safety cap is 3 rounds per criteria epoch and 9 rounds total. The crew-iterate guide explains confirmation gates, review convergence, and the optional bounded Claude /goal inner loop.

Safe by design

Write runs use isolated git worktrees. Enforced read-only reviewers can inspect an existing checkout; agy reviews use disposable, non-mergeable snapshot worktrees because Antigravity cannot enforce read-only access. The captain asks before merging or discarding. Squash is the default merge strategy; preserving an intentional commit stack is also supported.

Workers do not receive the captain tool surface. Claude Code and Codex workers get only a scoped send_message tool for returning structured findings.

Hosts and workers

| Integration | Captain host | Worker | Review mode | | --- | --- | --- | --- | | Claude Code | Global or project | Yes | Advisory read-only | | Codex | Global or project | Yes | Read-only | | Antigravity (agy, Gemini models) | Project only | Yes | Disposable snapshot | | Ollama / LM Studio | No | Text-only | No repo access | | OpenAI-compatible / generic | No | Yes | Depends on adapter |

Provider model pins are exact-or-refuse: Crew discovers provider-native model choices and rejects unknown names before allocating work. crew-mcp config can save one validated default for each built-in provider; per-call pins still win, and clearing a saved value restores the provider CLI default. Local OpenAI-compatible agents are useful for brainstorming and prose today, but they cannot read the repository until context injection is implemented.

Tool surface

| Stage | Tools | | --- | --- | | Dispatch | run_agent · continue_run · run_panel | | Observe | get_run_status · get_panel_status · list_runs · list_agents · list_models | | Decide | merge_run · discard_run · cancel_run · aggregate_panel | | Contract | create_criteria · confirm_criteria · get_criteria · revise_criteria | | PR watch | start_pr_watch · list_pr_watches · get_pr_watch_status · rearm_pr_watch · cancel_pr_watch · authorize_pr_watch_actions | | Preferences | get_crew_preferences | | Messaging | check_captain_inbox · acknowledge_messages — workers get only send_message |

That is 25 captain tools plus one worker-only tool. The tool reference owns the complete schemas and envelope contracts.

Extended documentation

• Documentation home
• Install and upgrade
• Agents and models
• crew-iterate
• Configuration
• Operations and troubleshooting
• Runtime architecture
• Durable PR watch

License

MIT