Orchestrator

What Orchestrator MCP server does

The Orchestrator MCP server connects an MCP host to coding-agent CLIs already installed and authenticated on the same machine. It can route consultations to Codex, Claude Code, OpenCode, GitHub Copilot, or the experimental Antigravity CLI. The host sends a request through orchestrator_consult, and the selected CLI returns a structured answer.

The main use case is getting another coding agent's view of a prompt, diff, or file context without opening a second terminal and manually copying the conversation. Follow-up requests can reuse a consultation_id, keeping the exchange associated with the same client conversation.

How it works

The Orchestrator MCP server runs locally and launches the selected vendor CLI as a subprocess. Authentication stays within each vendor's normal CLI login; the server does not read or store those credentials, request provider API keys, proxy provider traffic, or silently change models.

A Claude Code host can consult Codex, while a Codex host can consult Claude Code. Either can also route to OpenCode, GitHub Copilot, or Antigravity when those runtimes are configured. OpenCode uses hosted providers and does not run a model locally.

Ordinary consultations exclude the host's complete runtime. For example, Claude Code cannot consult Claude Code through the standard consultation path, and Codex cannot consult Codex. Review and workflow modes use a narrower runtime-and-model identity, allowing a same-runtime request only when the configured model is demonstrably different and versioned. The host cannot route work back to its own execution identity.

Setup and configuration

Install the package with Homebrew, uv, pipx, or pip; Python 3.11 or newer is required for the Python installation paths. Homebrew provides a prebuilt Apple Silicon package, while Intel macOS and Linux may build dependencies from source. The temporary uvx approach is also available in the project documentation.

Sign in to the agent CLIs you intend to use with their normal commands, such as codex login or claude auth login. OpenCode requires opencode auth login. GitHub Copilot must be logged in under the COPILOT_HOME directory used by the server and requires a Copilot plan.

Run orchestrator-mcp-server init --host claude or use codex for a Codex host. The command discovers supported local CLIs and writes ~/.orchestrator-mcp/config.yaml with mode 0600; it does not overwrite an existing file. Use --path to choose another location. The host then needs the configuration path and runtime environment settings when registering the MCP server.

Claude Code can install the provided plugin and run its setup command, which creates the configuration and runs a readiness check. If the server was previously added manually, remove the older entry to avoid running two copies.

Tools and capabilities

The documented consultation capability is orchestrator_consult. It returns a structured response from the selected coding agent and supports consultation_id for continuing a prior exchange. The project also describes reviewer panels, workflows, delegated edits in disposable worktrees, local records, cost estimates, and a dashboard, but the supplied material does not enumerate their individual tool names or full parameters.

Limitations and notes

Consultation calls block the turn that starts them and can take minutes. A background subagent can make the call if the host should continue working. Requests may include prompts, diffs, and file contents, which are sent to the selected provider under the user's own account and subject to that provider's billing and privacy terms.

The server depends on the relevant CLI being installed, authenticated, and configured. It does not provide local model execution for OpenCode, and it does not replace the authentication flow of any connected vendor. Antigravity support is identified as experimental. Google models are reached through Antigravity in the supplied documentation rather than through the Gemini CLI.