Extract SDK: PDF extraction for AI in your own code
Read every table, heading, figure, scan, pageon your own machine.
A library that turns a document into text in reading order, tables, headings, figures and chunks a model can use. It runs inside your own software.
Three steps.
Then it is just your code.
1# Python2pip install pdfluent34# Node.js5npm install @pdfluent/node67# Rust8cargo add pdfluent
Tables that stay tables.
And the rest of the page too.
Tables, cell by cell
Figures, found
Scans, read
More than PDF
Fast on one CPU.
Speed, quality fast
* Text extraction (quality fast) in SDK 1.2.0: 165 pages per second on one CPU against 85 for 1.1.0, measured in one run on one c3-standard-8 over 40 public arXiv papers (1,050 pages), 21 October 2026. Scientific papers only; no statement about scans or forms.
* Markdown extraction (quality fast) in SDK 1.2.0: 61.4 against 45.2 pages per second on one CPU, same run and machine. The Markdown output differs between the two versions on 20 of the 40 documents (a few hundred bytes), because 1.2.0 also contains extraction changes; the timings therefore compare slightly different output.
Ahead on quality.
Quality, quality auto
- PDFluent 1.4.00.9282
- opendataloader [hybrid]0.907
- nutrient0.885
- docling0.882
- marker0.861
- unstructured [hi_res]0.841
- edgeparse0.837
- opendataloader0.831
- mineru0.831
- pymupdf4llm0.732
- unstructured0.686
- markitdown0.589
- liteparse0.576
PDFluent 1.4.0 with quality auto, page headers and footers kept. Other engines as published in the benchmark README on 2 November 2026. Footnote
* Overall score 0.9282 (SDK 1.4.0, quality="auto", keep_furniture=True, layout, table and formula models installed) on all 200 documents, with the benchmark’s own harness and evaluator, measured 11 November 2026 on the final macOS arm64 release wheel. Competitor figures are taken from the benchmark’s public README as read on 3 November 2026 (highest overall there: 0.907, opendataloader [hybrid]), not re-run by us. The leaderboard engines ran on their own setups while PDFluent was measured on macOS arm64, and table structure depends on the platform. PDFluent is not on that leaderboard. Other engines have submissions open that are not yet on the leaderboard. One run, no confidence interval. The PDFluent row keeps page headers and footers because the benchmark’s ground truth counts them as content. PyMuPDF and MuPDF are trademarks of Artifex Software, Inc. PDFluent is an independent product of Innovation Trigger B.V. and is not affiliated with or endorsed by Artifex.
The small print,
up front.
Would you rather not run anything? Extract Hosted is coming soon: the same engine behind one request.
- Licence
- Free to use under the AGPL-3.0. A commercial licence is the alternative if you do not want to publish your own source.
- Languages
- Rust, Python, Node.js, WebAssembly, Java and .NET, plus a command-line tool.
- Element ids
- In Python, every block carries an element_id, and the same input gives the same ids.
- Agents
- pdfluent-mcp is an MCP server for Claude Desktop, Claude Code, Cursor, VS Code and Codex. It runs over stdio on your machine; your documents are not uploaded. Set it up
- From PyMuPDF
- A PyMuPDF-compatible layer is available as a Preview.*
* The PyMuPDF-compatible layer is a Preview in SDK 1.4.0. PyMuPDF and MuPDF are trademarks of Artifex Software, Inc. PDFluent is an independent product of Innovation Trigger B.V. and is not affiliated with or endorsed by Artifex.
Four questions.
What does it cost?+
Which tier should I use?+
Can my agent use it?+
What if I do not want to install anything?+
Try it on a
document you know.
Install it, open one file you can judge, and read what comes back.
Need to edit PDFs instead? PDF API