anydoc makes office files locally readable for AI agents

What this is about

anydoc is an open-source Rust library from Firecrawl that converts office and document files into GitHub-Flavored Markdown. Supported inputs include Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, and text-based PDFs. Search systems, retrieval pipelines, or AI agents can then process the result.

The tool is not a language model and does not write summaries itself. Its value appears one step earlier: it attempts to transfer headings, lists, tables, footnotes, equations, and other structures from different file types into one text format. anydoc is available through a command line, Node.js, Python, Rust, or WebAssembly and is published under the MIT license.

What anydoc actually does

anydoc tries to detect formats from file contents instead of relying only on extensions. Internally, it maps every supported format into one common document model and renders Markdown from that model. Similar structures should therefore produce similar output regardless of the source format. Embedded images remain available as binary data in the document model, while their alternative text appears in Markdown. Mathematical equations are emitted as LaTeX where possible.

The browser demo runs locally through WebAssembly. The libraries and command line also process ordinary text-based documents locally. OCR is an important exception: anydoc alone cannot read image-only or scanned PDF pages. When a user enables the optional hosted OCR setting, the complete affected document is sent to Firecrawl Parse. That boundary matters for privacy and compliance.

anydoc also ships as an Agent Skill that teaches coding agents to convert documents through the command line before processing them. This makes conversion more reproducible than asking a model to improvise its way through a binary file.

Why it matters

Many AI workflows start with a messy collection of presentations, spreadsheets, old Word files, and PDFs. If every format requires a separate parser, maintenance, failure modes, and security exposure all grow. One local tool for this preprocessing stage can simplify the pipeline and provide inspectable intermediate files.

Firecrawl reports a benchmark covering 100 real documents and 14 formats. According to the vendor, anydoc reached a median conversion time of 4.4 milliseconds and a quality score of 81. However, the evaluation used Claude Sonnet 5 as the judge, the document corpus is not public, and the compared tools covered different format sets. The figures are useful evidence, but not independent proof.

In plain language

Imagine eight boxes labeled Word, Excel, PDF, and other formats. An AI agent wants to sort their contents but cannot reliably open every kind of packaging. anydoc works like a packing station that moves the contents into identical, labeled containers. If a box contains only a photograph with no readable text, however, the station needs an external reading service.

A practical example

A consulting team receives 60 project files: 20 Word documents, 15 presentations, ten spreadsheets, ten text-based PDFs, and five scanned PDFs. It first processes 55 files locally with anydoc and stores the Markdown in an isolated workspace. A script then checks whether headings, tables, and source filenames survived.

The five scans are not uploaded automatically. The team first determines whether they contain personal or confidential information. Only approved files go to the optional OCR service; the rest use an internal OCR system. The team then indexes the reviewed Markdown for a research agent and links every answer back to its source file.

Scope and limits

First, anydoc does not provide local OCR for scanned pages; the hosted option changes the data flow materially. Second, successful conversion does not guarantee faithful content. Complex tables, layouts, embedded objects, and equations should be sampled against the original. Third, Markdown is flatter than many office formats: comments, revision histories, macros, or visual relationships may be lost or unsuitable for the target task.

Unknown files should also be processed in a constrained environment. Parsers handle complex binary formats and belong in the same security process as other import components. The sensible first test uses a representative, non-confidential document set and explicit quality criteria.

SEO & GEO keywords

anydoc, Firecrawl, Document AI, Markdown conversion, Agent Skill, local document processing, Word to Markdown, PDF to Markdown, WebAssembly, retrieval-augmented generation, OCR, open source