Grok (xAI API)

  • Connects:
    • Request: a remote MCP server is an entry in the request's tools array: { "type": "mcp", "server_url", "server_label" }. Optional fields are server_description, allowed_tools, authorization (sent as the Authorization header) and headers. (xAI: remote MCP)
    • Hosted only: only Streamable HTTP and SSE work, and xAI's servers make the connection, so only the hosted log works. logMcp on a laptop doesn't.
    • Not supported: require_approval and connector_id from OpenAI's Responses API.
    • Default tools: with no allowed_tools, every tool log exposes goes into the model's context. Narrowing it to the tools a job needs keeps that small.
  • Draws MCP Apps: xAI's page doesn't mention MCP Apps, resources, prompts or server instructions, and Grok isn't in the support matrix.
    • Views: report comes back as text.
    • Sign-in: happens in the log console.
  • Guidance: there are no prompt hooks and no confirmed instructions, so the model routes each request with getGuide({ ask: "<the owner's words>" }), as getGuide's description says. Keep getGuide in allowed_tools. server_description is the one field the API offers for describing log to the model, and it should say the same.
  • The Grok app: whether grok.com takes custom MCP connectors isn't documented on xAI's page.

On this page

No Headings