Summary
Long PDFs, not screenshots — that is where the 1000-page ceiling points.
Results land on disk as timestamped JSON rather than only in the conversation, which is what makes it usable for a batch: run the documents through, then process the files afterwards. The two tools differ in one thing worth knowing before you call the URL one — it will not guess the file type, you have to say image or pdf.
What it is
A server over Mistral AI's OCR API that processes a local file or a URL and writes the result to disk.
What you get
- process_local_file takes a filename from the directory you configured
- process_url_file takes a URL, with the file type stated explicitly as image or pdf
- JPG, JPEG, PNG, GIF and WebP images, plus PDF and other document formats Mistral OCR handles
- Results written as JSON into an output directory, named with a YYYYMMDD_HHMMSS timestamp
- A Docker image, with the local file directory mounted in
Requirements
MISTRAL_API_KEY, and OCR_DIR pointing at the directory holding the files to process — inside the container that path is always /data/ocr. Run it through Docker with the directory mounted, or locally with uv. Mistral enforces the limits: files up to 50MB and documents up to 1000 pages.