<- all tokdocs

anydoc: Firecrawl's Rust Converter Turns Office Files Into LLM-Ready Markdown in Under 5ms

Watch on TikTok

View on TikTok ->

Author: chase_ai_

Firecrawl released anydoc this week, a tool that converts Word docs, PowerPoints, Excel sheets, and PDFs into clean Markdown. The headline number is speed: a median conversion time of 4.4 milliseconds per document, versus 52.5ms for the next-fastest tool in Firecrawl's own benchmark. It ships as an MIT-licensed library you can run locally or call as an API, and it installs as an agent skill so a coding agent can read any document it encounters.

What anydoc Actually Does

anydoc is a Rust library that takes a typical office file and outputs GitHub-Flavored Markdown. The supported formats span Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, and PDF. The point is to produce LLM-ready text: instead of feeding a model a binary .docx or a messy PDF, you hand it clean Markdown with headings, lists, and tables preserved.

The README notes it includes bindings for Node.js, Python, and the browser. The browser demo runs the library as WebAssembly, so files are converted on your own machine rather than uploaded to a server. That matters for anyone converting sensitive documents.

The Benchmark Numbers

anydoc was measured against six other converters on 100 real-world documents. The scoring runs 0 to 100 where higher is better, and speed is the median time to convert one document. anydoc leads across the board in Firecrawl's table.

Benchmark table comparing anydoc against six other document converters

The comparison, as shown in the video:

Tool Formats Median ms Score Completeness Structure
anydoc 14/14 4.4 81 87 79
libreoffice 12/14 1129.5 40 59 42
unstructured 8/14 572.9 63 76 59
markitdown 6/14 134.8 65 78 66
pandoc 5/14 102.1 56 74 57
docling 4/14 513.6 57 60 60
mammoth 1/14 52.5 70 84 71

Two things stand out. First, anydoc covers 14 of 14 formats while the next-best covers 12. Second, the speed gap is not incremental. At 4.4ms it is over 20x faster than mammoth, the closest competitor on time, and hundreds of times faster than libreoffice or docling. The quality scores are AI-judged, so treat them as directional rather than absolute, but anydoc tops every column in this table.

Why It Runs This Fast

The speed comes from the architecture. anydoc is pure Rust with no ML models and no external service calls for the core conversion. Every input format parses into one shared document model and renders through one Markdown serializer. That means a .doc from 2003 and a fresh .pptx come out consistent, and there is no network round trip or model inference slowing things down.

Firecrawl says anydoc powers Firecrawl Parse. If you would rather not self-host, the hosted API gives you the same conversion plus OCR models for scanned pages. So the local library handles the fast structured path, and the paid API adds OCR for images and scans where you need it.

The Agent Skill Angle

anydoc ships as an agent skill, installable with a single command so your agent can read any document it runs into. This is the practical hook. If you are building an agent that ingests user-uploaded files, you drop in anydoc and the agent can parse a spreadsheet or a slide deck without a separate parsing pipeline. The README shows the install as npx skills add firecrawl/anydoc.

Key Takeaways

  • Speed is the differentiator. 4.4ms median conversion beats the next tool by more than 10x and makes anydoc practical for high-volume ingestion.
  • Format coverage is complete. 14 of 14 tested formats, ahead of every competitor in Firecrawl's benchmark.
  • It runs locally or as an API. MIT-licensed Rust library with Node.js, Python, and browser bindings, plus a hosted Firecrawl Parse API that adds OCR.
  • Treat the quality scores as directional. The completeness and structure numbers are AI-judged and come from Firecrawl's own benchmark, so validate on your own documents before committing.
  • It plugs into agents. Ships as an installable agent skill so a coding agent can read documents it encounters.

Resources

  • anydoc on GitHub - The MIT-licensed Rust library, with Node.js and Python bindings and benchmark details.
  • anydoc browser demo - Try the WebAssembly demo that converts files locally in your browser.

Published August 6, 2026. Writeup generated from a favorited TikTok.