Scan to Markdown: Free Local OCR for PDFs and Images (Nothing Uploaded)
Convert scanned PDFs, photos, and screenshots to Markdown with free OCR that runs entirely in your browser. How local OCR works, where it shines, where it fails, and when to ask for server-side OCR.
Most "PDF to Markdown" converters silently fail on scanned documents — and most online OCR tools solve that by uploading your scan to someone else's server. There is a third option: run the recognition on your own device, in the browser, and keep the file private end to end. That is exactly what the free OCR to Markdown tool does.
Why scans need OCR at all
A text-based PDF contains actual characters: select the text, copy it, done. A scanned PDF (or a photo of a page) contains only pixels shaped like letters. No converter can extract characters that aren't there — something has to read the image first. That something is optical character recognition.
How local OCR works
The tool loads a compact recognition model into your browser on first use (this one download needs a connection), then processes everything on-device:
- Drop a scanned PDF, photo, or screenshot — up to 5 files, 10 pages per PDF.
- Pick the document's language. This matters more than people expect: the recognizer compares shapes against language-specific training data.
- Watch the progress bar while each page is read, then get Markdown — not a raw text dump, but paragraphs reflowed, wrapped lines joined, headings and lists detected.
Close the tab and nothing remains. No account, no upload queue, no retention policy to trust, because there is nothing to retain.
Where local OCR shines
- Clean scans at 200–300 DPI with straight, printed, high-contrast text. This is the bread and butter: contracts, statements, book pages, printed forms.
- Sensitive documents. Medical records, legal files, internal memos — the privacy story isn't a policy, it's physics. The bytes never leave the machine.
- Quick jobs. A two-page scan on a modern laptop converts in well under a minute with no signup flow in the way.
Where it honestly falls short
No small model reads everything, and you should know the failure modes before trusting the output:
- Low-resolution or skewed photos. Blurry, tilted, shadow-covered phone pictures are the number one cause of garbage output. Fill the frame, hold steady, boost contrast.
- Handwriting. Cursive and most handwriting defeat compact models. Print is fine; script is not.
- Dense tables and tiny type. OCR reads characters, not structure — verify every cell of a recognized table, and proofread numbers character by character (0/O, 1/l, 5/S are classic confusions).
- Decorative and unusual fonts. Anything far from standard book/print faces degrades accuracy.
- Speed on big jobs. Your device does the work, so a 50-page archive on an old phone is a long coffee break. The in-browser tool caps PDFs at 10 pages for exactly this reason.
The rule is simple: always proofread OCR output before publishing it or feeding it into an AI pipeline. OCR errors are confident-looking — a misread name or figure won't announce itself.
Practical tips for the best results
- Match the document language first, convert second.
- Scan at 200–300 DPI; for photos, use good light and shoot straight down.
- Straighten and crop to the page before uploading — most scanning apps do this in one tap.
- Keep "join wrapped lines" on for prose so visual line breaks don't fragment your paragraphs.
- For tables, convert, then check cell by cell against the scan.
When to step up to server-side OCR
Local OCR covers clean scans beautifully. For the hard cases — handwriting, poor-quality archival scans, dense tables, whole multi-hundred-page documents, or steady high volumes — larger models on dedicated hardware do meaningfully better. That option exists and access is currently granted on request: describe your material on the contact page under "Server-side OCR access" and you'll get an answer from a human.
The bottom line
OCR doesn't have to mean uploading sensitive scans to a stranger's server. For clean, printed material, a free local pass plus two minutes of proofreading gets you publish-ready Markdown — privately. For everything harder, know exactly where the boundary is, and ask for the bigger engine instead of fighting the small one.