Skip to content
mistral.rs

Expose mistralrs as an MCP server

mistral.rs can expose the loaded model as an MCP (Model Context Protocol) server: a chat tool over JSON-RPC 2.0 that any MCP client can call.

Terminal window
mistralrs serve -m Qwen/Qwen3-4B --mcp-port 4321

--mcp-port starts an additional listener. The port rules:

  • It shares --host with the main HTTP API.
  • The OpenAI-compatible API on --port (default 1234) still runs alongside.
  • --mcp-port must differ from --port.
  • The bind is validated at startup, so failures surface before serving.

In a TOML config, the equivalent is mcp_port under [server]:

command = "serve"


[server]
port = 1234
mcp_port = 4321

Clients connect to http://<host>:<mcp_port>/mcp. Each call is a POST /mcp with a JSON-RPC 2.0 body.

Methods

Section titled “Methods”
  • initialize: returns {"capabilities":{"tools":{}},"instructions":...,"protocolVersion":"2025-11-25","serverInfo":{"name":"mistralrs","version":...}}.
  • ping: returns {}.
  • tools/list: returns the chat tool. The list is empty if the loaded model does not have text input and output modalities.
  • tools/call: runs the chat tool.

Anything else returns JSON-RPC error -32601 (method not found). A body with jsonrpc other than "2.0" returns -32600; tool execution failures return -32603.

The chat tool

Section titled “The chat tool”

You can pass any OpenAI ChatCompletionRequest field in arguments; the advertised schema only documents the common ones. The schema:

  • Requires messages: an array of {role, content} objects with roles user, assistant, or system.
  • Documents max_tokens and temperature.
  • Defaults model to "default".
Terminal window
curl http://localhost:4321/mcp \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/call",
    "params": {
      "name": "chat",
      "arguments": {
        "messages": [{"role": "user", "content": "Hello!"}],
        "max_tokens": 50
      }
    }
  }'

The result is MCP tool-call content:

{"content": [{"type": "text", "text": "Hello! How can I help?"}]}

Authentication

Section titled “Authentication”

The MCP endpoint has no built-in authentication. For non-localhost use, place an authenticating proxy in front.

See also

Section titled “See also”