OCR a scanned contract without uploading it

What this tool does, exactly

A signed agreement is the document people are least willing to hand to a website, and the one they most often need as text: to search it, quote a clause, or find the date buried on page nine. This page reads it here — every page drawn and read in this browser tab — and gives back the text, the text as Markdown with its clause numbering, and the scan itself made searchable. Nothing about it is sent anywhere.

Why it stays here

A contract carries names, sums, signatures and often a confidentiality clause of its own; a document under legal privilege or a non-disclosure agreement is exactly the kind you may not be allowed to hand to a third party. There is no server behind this page and no request that carries a file: the network panel shows the reader's own code arriving from this site, and nothing carrying your file leaving. That is a fact about the mechanism, not legal advice about what you may disclose.

What comes back

Three files. The text in reading order, for searching and quoting. A Markdown version in which each line is kept as read, so a clause number stays at the start of its line and 7.2 is still findable as 7.2. And the scan with an invisible text layer laid over the original pages, which opens in any viewer and can be searched, with the layer read back before the file is offered.

Signatures, stamps and handwriting

The reader reads print. A signature is not text and becomes nothing, or a short doubtful line the page lists by name. A stamp over a paragraph was not measured; expect the lines under it to read worse and to appear in the doubtful-lines list. Handwriting was not measured and is not promised — check any handwritten addition against the scan.

Tables and schedules

The fast reader gives a table as lines of cells in reading order. The Deep reader behind the Pro door reads a page as a page, so a fee schedule comes back as rows of cells in the Markdown. It takes about twenty seconds a page on a laptop's graphics chip and is free during the preview.

What it was measured at

0.23 % of characters wrong on a synthetic 200 dpi scan of a printed English page, 0.71 % in Portuguese, 0.64 % in Russian with the second recogniser, in this site's own tests. The one systematic slip on English print: an old-style zero — the kind that sits low like an o — is read as the letter, which matters in account numbers. Check those.