Consult MCP

What consult-mcp does

The consult-mcp MCP server connects an MCP host to locally installed advisor CLIs: Claude Code, Codex, Grok, and Google's Antigravity CLI. An agent can ask one advisor a question, send the same question to several advisors in parallel, or continue a previous conversation by providing its session ID. Advisors can read the selected working directory, which normally defaults to the repository where the host started the server.

The server is intended for the 🤖 Coding Agents category. Its MCP instructions encourage agents to seek another opinion after failed fixes, before hard-to-reverse technical decisions, and before declaring a change complete. Advisors are instructed to use the repository as context while treating each question as self-contained.

How it works

Each advisor runs through the CLI already installed and signed in on the machine. The selected model and reasoning effort can be passed through to that CLI; otherwise, its configured default is used. The result includes the advisor identity, answer, session ID, optional model, and elapsed time.

The consult-mcp MCP server can collect a Git diff and send it to multiple advisors for review. Without a base branch, the review covers staged and unstaged changes against HEAD and identifies untracked files. With a base such as main, it reviews the branch from its merge base while including uncommitted work. Review requests can also specify a focus area. Advisors report concrete issues with severity, file and line references, scenarios, and suggested fixes, while remaining read-only.

Optional live-session features can discover running sessions, ask them questions, and send messages. Task dispatch can assign coding work to Codex or Grok workers, then report status or cancel the task.

Setup and configuration

Node.js 22 or newer is required on macOS, Linux, or Windows. At least one of the claude, codex, or grok CLIs must be installed and signed in. Run npx -y consult-mcp@latest install to register consult with agent CLIs found on PATH, then restart existing agent sessions. npx -y consult-mcp@latest doctor checks CLI versions, registration, settings, hooks, and Antigravity's read-only configuration. Add --ask to verify that each advisor can answer.

Installation updates user-level MCP configuration and, when session messages are enabled, adds consult hooks and instruction snippets to supported CLI files. Existing unrelated settings and hooks are preserved. npx -y consult-mcp@latest uninstall removes the registration. A global npm installation or a checkout-based node dist/cli.js install can be used when pinning a specific build is preferred.

Tools and capabilities

The consult-mcp MCP server exposes these capabilities when enabled:

• ask_agent for one advisor, with model, effort, session, and working-directory options.
• ask_agents for parallel answers from several advisors.
• review_changes for multi-advisor Git reviews.
• list_agents for installed CLIs, versions, and available model IDs.
• list_sessions and ask_session for live sessions.
• send_message for session messaging.
• dispatch_task, task_status, and cancel_task for task workers.

Limitations and notes

Antigravity is an advisor only; consult does not register itself with that CLI because it must remain read-only without MCP servers enabled. On Windows, live-session discovery and session messaging are unavailable because they depend on POSIX tools and shell hooks. Other documented functions work there.

Advisor requests run through command-line interfaces and use those CLIs' existing subscriptions. The server does not require API keys. Review input is limited by operating-system command-line length, so advisors may receive a truncated diff and are told to read changed files for the remainder. Review collection disables external Git diff helpers, text conversion, filesystem monitors, and clean/process filters.