Turn a flat PDF back into clean, compiling LaTeX.
Uncompiler reconstructs research papers into editable, multi-file LaTeX projects — idiomatic code, a rebuilt bibliography, real \cite keys — and verifies the project compiles before you ever see it.
Drop a PDF here
or click to choose · up to 40 pages during beta · free
- ✓ Verified compile on every job
- ✓ Fidelity report included
- ✓ Unconverted parts flagged, never hidden
- ✓ Files deleted after 30 days
Lost the source?
The .tex is gone, the deadline isn't. Get an editable project back instead of retyping.
Building on a legacy paper?
Turn a 2004 scan into source you can actually extend, cite, and recompile.
Migrating an archive?
Batch through the API. Every document returns with its own fidelity report.
How it works
Four stages, one promise: what comes out compiles.
Structural parse
A vision model reads your PDF the way a typesetter would: sections, reading order across columns, display math, table geometry, figures and their captions.
Idiomatic reconstruction
The document becomes a modular project — main.tex, sections/, macros.tex — written the way you'd write it: \binom not nested arrays, booktabs tables, amsmath environments.
Citations, verified
Reference strings are matched against CrossRef. Verified entries land in references.bib and every inline [12] becomes a real \cite{key}. Anything unverified is kept verbatim and flagged — never silently wrong.
Sandboxed compile loop
The project compiles in an isolated sandbox. Failures are diagnosed from the log and fixed, up to a bounded number of attempts. It ships only when it builds — or it's free.
What lands on your disk
Not a text dump — a project. It opens in Overleaf, Prism, or your editor and compiles on the first try. The bibliography is real BibTeX verified against CrossRef, and the preview PDF in the zip is the proof.
And when something couldn't be converted — a hand-drawn diagram, an unverifiable reference — it's written down in WARNINGS.md, not hidden and hoped for.
Fidelity, measured honestly
We optimize for content parity, not pixel parity. Every job ships a report card: text similarity against your original, element-by-element counts for equations, tables, figures and citations, and a compile record including how many fix attempts it took.
Green
Everything reconstructed and verified. Compile clean.
Yellow
Compiles, with flagged items — e.g. a figure kept as an image, or an unverified reference.
Red
Didn't reach a clean compile. You get the partial project, the log — and you don't pay.
Built for humans and agents
Everything on this page is available as a REST API and an MCP server, so your scripts — or your AI agents — can uncompile documents without touching a browser. Create an API key in the dashboard.
# REST
curl -X POST https://uncompiler.com/api/v1/jobs \ -H "Authorization: Bearer uk_..." \ -F "file=@paper.pdf" curl https://uncompiler.com/api/v1/jobs/JOB_ID/result \ -H "Authorization: Bearer uk_..." -o paper-latex.zip
# MCP (Claude Code, etc.)
claude mcp add uncompiler \ https://uncompiler.com/api/mcp/mcp \ --header "Authorization: Bearer uk_..." # tools: uncompile_pdf, job_status, list_jobs
Questions, answered straight
Is this just OCR?
No. OCR gives you text; extraction tools give you one long file you spend an afternoon repairing. Uncompiler delivers a project: modular files, a rebuilt bibliography with real \cite keys, and a compile that's been verified before you download it.
How is this different from Mathpix and other converters?
Those tools are genuinely good at extraction, and we're not shy about saying so. The difference is what lands on your disk: a flat export you clean up versus a structured project that already compiles, plus a report that tells you exactly what didn't convert. We compete on the last mile, not the first.
Will the output look identical to my PDF?
No — and that's deliberate. LaTeX separates content from styling, so we target content parity: every paragraph, equation, table and citation, in order, in clean source. The compiled preview uses a modern template; swap the class file to restyle it in seconds. Chasing pixel-identical output produces worse code, and you're here for the code.
What about scanned PDFs?
Supported, and honestly the hardest mode. During beta, scans run free precisely because we're calibrating on them — expect more yellow flags in the fidelity report, and every one will be listed.
What happens to my files?
Processing happens in isolated sandboxes. Files are deleted 30 days after your job finishes, sooner on request. Your documents are never used to train models.
What will it cost after the beta?
Per paper — around $19 — with a simple rule: if your project doesn't compile, you don't pay. Beta conversions are free in exchange for feedback.
Get your paper back.
The beta is free and capacity-limited. Leave your email and we'll open a slot — usually within a day.