Firecrawl ships open-source anydoc for faster AI document ingestion

Why it matters

Anydoc extends Firecrawl from web extraction into the files that feed RAG and agent systems. Open source drives adoption, while the hosted /parse endpoint captures paid workloads.

Firecrawl, founded by Caleb Peffer (@CalebPeffer), Eric Ciarla (@ericciarla) and Nicolas Silberstein Camara (@nickscamara_), launched anydoc in a post on X on August 4th, pitching the open-source Rust engine as a faster way to turn business documents into Markdown for AI systems.

The San Francisco founders have moved steadily upstream through the AI data stack. They previously built Mendable, an AI chat product for documentation, and SideGuide, a developer-learning tool. Peffer wrote in Firecrawl's 2025 funding announcement that Mendable repeatedly ran into the same infrastructure problem: acquiring clean, structured web data required each AI developer to assemble scrapers, proxies and post-processing systems. Firecrawl grew from the founders' decision to sell that data layer directly.

Anydoc extends that thesis from websites into files. Firecrawl says the engine converts PDFs, Word documents, slide decks and 10 other formats into Markdown. The company claims parsing latency below roughly five milliseconds and says anydoc is fully open source. Firecrawl has not supplied benchmark methodology with the announcement, so the latency figure remains a company-reported performance claim rather than an independently tested comparison.

The release also clarifies what sits underneath Firecrawl's hosted document product. The company introduced its /parse endpoint on April 28th for developers uploading local or non-public documents. That API previously documented support for PDF, DOCX, DOC, ODT, RTF, XLSX, XLS and HTML files up to 50 MB, returning Markdown, structured JSON or summaries while preserving reading order and tables.

Firecrawl already described /parse as using a Rust-based engine that was up to five times faster than its prior pipeline. Anydoc gives that engine a standalone identity, adds slide decks to the formats Firecrawl is publicly promoting, and turns the parsing layer into an open-source distribution channel for the paid API.

The five-millisecond claim has limits

Document parsing does not follow one uniform execution path. Text embedded cleanly in an Office file or PDF can be extracted directly. Scanned pages, handwritten material and complex layouts can require optical character recognition and layout analysis, which carry substantially different processing costs.

Firecrawl's earlier Fire-PDF documentation described a system that classifies individual PDF pages before selecting direct extraction or a heavier layout-aware path. The company reported average processing below 400 milliseconds per page for that broader PDF pipeline. Its /parse documentation separately warns that image-only PDFs go through OCR and that results depend on scan quality. The new sub-five-millisecond figure therefore should not be read as end-to-end latency for every supported document or OCR workload.

That distinction matters for developers building retrieval-augmented generation, document search and agent workflows. The parser's speed on clean files affects ingestion throughput, while performance on tables, scans and irregular reading order determines whether the resulting context is useful. A fast conversion that scrambles a financial table or slide sequence simply moves the failure downstream into retrieval and model output.

Firecrawl moves into a crowded parsing market

Anydoc enters a field that already includes local libraries and hosted services. Docling, an open-source project started at IBM Research, supports PDF, DOCX, PPTX, XLSX, HTML, images and other formats, with export to Markdown or JSON. Other developers can choose specialized document APIs or assemble OCR and conversion components themselves.

Firecrawl's advantage is distribution through a broader data platform. The same account and API surface can already search, scrape and crawl the web, interact with pages and parse uploaded files. By installing anydoc behind /parse, Firecrawl can use open source to win developers while selling hosted execution, security controls and managed infrastructure around the engine.

Firecrawl has capital to pursue that expansion. It raised a $14.5 million Series A led by Nexus Venture Partners in August 2025, with Y Combinator, Zapier, Tobias Lutke, Abhinav Asthana and Matt McClure participating. Firecrawl currently reports $16.2 million in total funding. The valuation was not disclosed.

Anydoc pushes Firecrawl beyond its original position as a web-scraping API. Peffer, Ciarla and Camara are betting that developers will prefer one ingestion layer for both live websites and the files that sit inside company systems. Open-sourcing the parser lowers the barrier to adoption; routing the same engine through /parse gives Firecrawl a path to convert that adoption into API usage.