Document Parsing API

Firecrawl Parse turns a document into clean markdown your agent can read. Upload a PDF, Word doc, spreadsheet, or HTML file and Parse returns the same structured markdown, JSON, or summary you'd get from scraping a web page. One endpoint, one consistent output format, no matter what goes in.

Nine extensions today: .pdf, .docx, .doc, .odt, .rtf, .xlsx, .xls, .html, and .htm. Anything else returns an UNSUPPORTED_FILE_TYPE error. Files can be up to 50 MB per request.

Teams building document ingestion for RAG, invoice and contract extraction, research pipelines over academic PDFs, and any agent that gets handed a file instead of a URL. Over 1.25M developers and 150,000+ companies build with Firecrawl, including teams at Zapier, Botpress, Vercel, and Coinbase.

Use Scrape when you have a URL and want the page behind it. Use Parse when you already have the bytes: a file upload from your users, a document pulled from S3, an email attachment. Both return the same clean markdown, so the rest of your pipeline doesn't have to care where the content came from.

Two open-source Rust engines do the heavy lifting. pdf-inspector reads a PDF's internals, its fonts, text operators, and image coverage, to classify every page in milliseconds without rendering it. anydoc converts Word, Excel, OpenDocument, and RTF files straight into GitHub-Flavored Markdown, with a median conversion time under 5 ms per document. Neither loads a model or calls an external service.

No, and that's the point. Parse classifies each page first and only sends the ones that actually need it to the GPU. Roughly half of real-world PDFs are text-based and never touch OCR, which is where most of the speedup comes from. You can override the behavior per request with the pdf parser mode: fast for text-only extraction, auto for text-first with OCR fallback, or ocr to force every page through OCR.

They survive. A layout model detects tables, formulas, text blocks, and headers as separate regions, then tunes extraction per region: tables get a higher token budget, formulas are preserved in LaTeX, and reading order is predicted for multi-column documents. Headings, nested lists, footnotes, and links come through as ordinary markdown.

Yes. Request the json format with a schema that describes one table row, and Parse returns the table as an array of typed row objects. To keep the original layout instead, turn on layout blocks for PDFs, and each table comes back as a table block with its HTML markup, page number, and bounding box, at no extra cost.

Markdown by default. You can also request summary, links, images, cleaned or raw HTML, and json with a schema or a prompt to pull typed fields straight out of the document. Request several at once and they all come back in the same response.

Set maxPages to bound the work and timeout to control how long you're willing to wait, up to 300 seconds. The response reports numPages alongside totalPages, so totalPages greater than numPages tells you the output was truncated. For batches of files, call Parse once per file in parallel; there's no batch upload variant.

Parse supports Zero Data Retention, and results are never cached, so every call re-parses from the bytes you sent. Firecrawl is SOC 2 Type II certified, and a DPA is available. You can also redact personally identifiable information from the returned markdown with redactPII.

1 credit per page, plus any LLM-backed formats you request such as json or summary. PDFs are billed per page. Because results are never cached, re-uploading the same file is billed again.

Yes. Free accounts include 1,000 credits per month, enough to parse a real stack of documents and try every format. No credit card required.

Yes. pdf-inspector and anydoc are both MIT licensed and published on GitHub, with bindings for Python, Node.js, and Rust, plus a WebAssembly build that runs in the browser. Run them on your own infrastructure, or use the hosted endpoint and get OCR, layout detection, and structured extraction on top.