It keeps the structure, not just the words
Most OCR returns a wall of text and leaves you to work out which numbers belonged to which row. This returns Markdown: a table in the image comes back as a table, headings stay headings, and a list stays a list.
That is the difference between something you can paste into a spreadsheet and something you have to retype.
The instruction box changes what you get
Left empty it transcribes everything. Filled in, it does what you asked — "only the invoice number and the total, as JSON", "describe the chart", "list the ingredients".
It reads the picture rather than just recognising characters, so questions about what is in the image work as well as requests to transcribe it.
Documents go to a server — bear that in mind
Reading a picture takes a vision model on a graphics card, so this tool uploads. The image goes browser-to-storage on a signed address good for five minutes, and both it and the transcription are gone within about three hours.
Worth saying because of what people photograph: if the page is a contract, a medical letter or anything with an account number on it, decide whether you want it leaving your device at all. Nothing here is read by a person or kept, but "sent to a server" is still true and only you know whether that matters for this particular page.
Frequently asked questions
Which languages does it read?
Chinese, Japanese, Korean, English and many others, including mixed on the same page. It is noticeably better at CJK than most OCR, which is usually where the free tools give up.
Will it read handwriting?
Clear handwriting, often. Hurried handwriting, unreliably. Check anything that matters — it produces a confident answer either way, and a wrong digit in a transcribed number looks exactly like a right one.
Can I copy the text straight out?
Yes — the transcription appears on the page as selectable text with a copy button, not as another picture. That is the whole point of the exercise.
Does it cost anything?
No, and no account is involved. There is a cap on how many pages one network reads per hour; a long document is best done a few pages at a time anyway, since accuracy is easier to check in small pieces.
Something wrong with this tool, or an idea for it? Tell us