The 8 tools divide into inference — chat, fill-in-the-middle, vision, OCR, transcription — and Workflows, where an execution is started and then polled until it reaches a terminal state. Model choice stays explicit: each tool names the models it accepts, and mistral_vision requires one with vision support. workflow_execute returns RUNNING rather than a result unless it is told to wait, so anything long is asynchronous by default.
An MCP server over Mistral's API, covering both the inference endpoints and Workflow executions
- mistral_chat completions with temperature, top_p, seed, reasoning effort and a response format for structured output, returning the assistant text plus token usage.
- codestral_fim writes the code between a prompt and a suffix, for editor autocomplete or an edit whose boundaries are already known.
- mistral_vision takes messages mixing text and image_url parts against a vision-capable model such as pixtral-large-latest.
- mistral_ocr turns a PDF or image into per-page markdown, with page selection, table format, header and footer extraction, confidence scores and bounding-box annotations.
- voxtral_transcribe transcribes audio from a public URL or an uploaded file id, with a language hint, diarization and timestamp granularities.
- Workflow control: start an execution and either wait or take an execution id back, poll its status, and signal or query a run that is still going.
A Mistral account and MISTRAL_API_KEY. OCR and transcription read from a public URL or a file id from the Files API, so local material has to be uploaded first.
One command plus a key — npx mistral-mcp, then supply credentials