# mistral-mcp

A self-hosted MCP server that exposes the full Mistral AI API -- chat, OCR, Voxtral audio transcription, Codestral code completion, vision, batch, and durable workflows -- to Claude Code, Cursor, Zed, Windsurf, and other MCP clients.

mistral-mcp exposes Mistral AI's API surface as MCP tools, resources, and prompts, so an MCP client can call `mistral_ocr` to pull structured text and bounding boxes from a PDF or image, `voxtral_transcribe` to transcribe audio with optional speaker diarization, `codestral_fim` for fill-in-the-middle code completion, or `workflow_*` tools for Temporal-backed durable multi-step execution with human-in-the-loop signals, all without leaving the agent's tool loop.

It is self-hosted with no mandatory SaaS proxy, uses a bring-your-own Mistral API key, and can route every call to a user's own OpenAI-compatible endpoint (vLLM, TGI, LiteLLM, or an internal gateway) instead of Mistral's cloud. A configurable cache for document processing has its own retention window, and a lean `core` profile limits which tools are exposed.

It does not expose fine-tuning or user management, and is scoped to French- and English-language prompts.

Website: https://github.com/Swih/mistral-mcp