Firecrawl's anydoc gains OCR, so scanned documents also turn into machine-readable text
The MIT-licensed Rust library converts 14 document formats to Markdown. With OCR added, scanned pages are in scope too, at a reported median of 190 milliseconds per page.
Firecrawl has added optical character recognition to anydoc, its open-source document converter. The library can now turn not only digital documents but scanned pages into text.
anydoc is a library written in Rust, and its job fits in one sentence: it takes a document and converts it to clean Markdown. The 14 supported formats include Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV and PDF. The output is not flat text; headings, links, lists, tables, footnotes and equations as LaTeX are preserved.
Why it matters: when language models and agents want to read a corporate archive, format is the first obstacle. A folder mixing PDFs, Word files and scans is not something a model can digest directly. anydoc handles that intermediate layer and produces Markdown, the format models work with most comfortably.
The speed figures the company gives: a median of 4.4 milliseconds for documents that need no OCR, and 190 milliseconds for scanned pages that do. Layout detection, table and formula extraction also work on the OCR side.
There are three ways to use it: from the command line (npx @firecrawl/anydoc document.docx -o output.md), as an npm, pip or Rust package, and as a WebAssembly build that runs in the browser. The library is MIT licensed, so commercial use is open. OCR runs as a hosted service through Firecrawl Parse; the company says no signup is required and that an API key only raises the limits. There is an important distinction here: anydoc does not perform OCR locally. It sends the document it detects as scanned; and only that one; to Firecrawl Parse; the rest of the conversion never leaves the machine. For a corporate archive, that distinction means knowing exactly which documents go out.
Local OCR is coming. Firecrawl engineer Abimael Martell has announced that local OCR is available in anydoc and pdf-inspector, with a fast classification engine deciding which pages go to OCR. That capability is not in a published release yet: the latest version on both npm and GitHub is 0.2.4, whose notes describe hosted OCR. Once local OCR ships, scanned pages will be readable without the document leaving the machine; the step that actually matters for corporate archives.
The source is on GitHub.
