markitdown

The problem it solves is familiar to anyone who has fed documents to a model: a PDF is not text inside, it is character coordinates on a page. Hand one over directly and the table falls apart, headings blend into captions, and column order becomes a matter of luck.

Microsoft's markitdown converts the document into markdown while keeping the structure: headings stay headings, tables stay tables, lists stay lists. The model then reads the document the way a person does rather than as a stream of words.

What it handles: PDF, Word, Excel, PowerPoint, HTML, EPub, ZIP archives, images (with text recognition), audio (with transcription) and YouTube links. There is an MCP server too, if you would rather hand the conversion to an agent than run it yourself.

Why it is in this catalogue. A hundred and seventy thousand stars is not hype; it is how often "let the model read this document" comes up. MIT licence, installs through pip, runs locally — the documents go nowhere, which matters when they are contracts or an export from your own database.