The everyday use is cross-checking: ask Grok the same question your current assistant just answered, or hand it an image the client you are in cannot see. Because the full messages array goes through, you control the system prompt rather than inheriting one. Worth knowing before you wire it in: every call is billed to your own Grok AI key, and the function_calling tool returns Grok's decision about which of your functions to call — your code still has to run it.
A TypeScript server that wraps the Grok AI API so any MCP client can send a request to Grok and get the answer back as a tool result.
- chat_completion — a full messages array sent to Grok, with model, temperature and max_tokens all overridable
- image_understanding — an image by URL or as base64, with a prompt, answered by a Grok vision model
- function_calling — a messages array plus your tool definitions, with tool_choice set to auto, required or none
- Defaults you do not have to set: grok-3-mini-beta for text and function calling, grok-2-vision-latest for images
A Grok AI API key in XAI_API_KEY, obtained from the x.ai console — usage bills to that account. Node.js v16 or higher, built from the repository and launched by path from your client config.
One command plus a key — npx -y grok-mcp, then supply credentials