Pull the text out of an image, in your own browser.

Optical character recognition looks at the shapes in a picture and decides which letters they most likely are. The engine here is Tesseract, a long-running open-source OCR project; tesseract.js is the build of it that runs inside a browser, which is why nothing has to be uploaded for it to work.

What comes back is text, not a document. Columns, tables, headings and captions are flattened into lines in the order the engine thought you would read them. Bold stays a word; a table becomes rows of words with the columns run together. If layout matters to you, this is the wrong shape of tool.

Resolution and contrast, mostly. Dark text on a plain light background at a decent size reads almost perfectly; the same words photographed small, at an angle, over a patterned background or in low light come back full of guesses.

If you are photographing a page rather than scanning it, get the page flat, fill the frame with it, and keep the camera square to it. Straightening a photograph before you feed it in does more for the result than anything the tool can do afterwards.

Digitising printed documents so they can be searched and edited rather than retyped. Getting the figures off a receipt or an invoice for expense records. Quoting a page from a book or a report without copying it out by hand. Lifting the text out of a screenshot, an error dialog or a scanned form where nothing is selectable.

The common thread is that the words already exist and somebody would otherwise be retyping them. That is worth a few seconds of a laptop’s attention; it is rarely worth a person’s afternoon.

The page loads the English language data and nothing else, so it reads English and the Latin alphabet generally. Feed it Tamil, Hindi or Arabic script and it will return confident nonsense rather than an error, which is worse than failing.

Tesseract itself has trained data for well over a hundred languages. This page does not offer a picker for them yet, and until it does it claims only the one language it actually loads.

We pull text out of screenshots and scans often enough to want a box for it, and we were not willing to upload a client’s paperwork to a free OCR site to get it. So the recognition runs in the browser, and since it costs us nothing to run, it is free and it stays free.