Handwriting recognition
Handwriting read the way you choose.
To search notes by content and turn them into text, the app reads what you wrote by hand. There are five engines: two run on your Mac with no internet and no keys, three are cloud services on your own key. Which one gets used is up to you.
Whatever is read once is stored and reused everywhere: in Spotlight, in AI Studio, in integrations and over MCP.
| Engine | Where it runs | What it needs |
|---|---|---|
| Apple Vision | this Mac | nothing — already in macOS |
| Tesseract | this Mac | a folder of language files |
| Google Cloud Vision | Google’s cloud | an API key |
| Yandex Vision OCR | Yandex’s cloud | an API key |
| AI model | at a provider | a connection from AI Studio |
Five engines
Who reads your handwriting
the green ones never go online
Apple Vision
this MacBuilt into macOS, asks for nothing. Good with Latin and Cyrillic script.
Tesseract
this MacAn open-source engine that runs on your Mac. It keeps languages as separate files — download the ones you need and point it at the folder. Useful for languages Apple doesn’t cover.
Google Cloud Vision and Yandex Vision OCR
cloudCloud services on your own key. Pages go to them as images. At Yandex, handwriting and printed text are read by different models, so the language lists differ.
AI model
cloudAny AI model from your AI Studio connections that is marked as able to do OCR. For each one you can allow printed and handwritten text separately — some models read scans beautifully but invent things on handwriting.
Which one reads better depends on your language, your handwriting and the page. We quote no percentages: they would have been measured on somebody else’s handwriting. Try two engines on a note you know well.
Handwriting and print
Two separate choices
a note and a scan are different jobs
A handwritten note and an imported scanned PDF are different jobs for recognition. So the engine and the language are chosen separately for handwriting and for printed text. Each menu lists only the engines that can read that kind of text, and the language list comes from the engine itself.
Spotlight picks no engine of its own: a note goes to the engine you chose for handwriting, a printed PDF to the one you chose for print.
What leaves the Mac
A cloud engine is switched on through a prompt
and off again with one click
Local engines ask for nothing
Apple Vision and Tesseract read on this Mac. Pages go nowhere, so there is nothing to agree to.
Cloud engines are off until you switch them on
Switching one on brings up a prompt: “Page images will leave this Mac and go to Google Cloud Vision with your key.” Cancel leaves the switch off.
Keys live in the macOS Keychain
Cloud service keys are kept where the rest of your passwords are, and they never leave the Mac. Privacy →
Read once
The result is shared across the app
and never read again
Stored with the document
The text that was read is kept alongside the document’s version and the engine’s name. Search, AI Studio, integrations and MCP take it from there instead of reading again.
A PDF with its own text isn’t recognized
If the file already has a text layer, that is what gets used. Blank notebook pages are skipped as well: out of 50 pages with 11 written on, 11 are read.
A model refused — the pass carries on
One document a model wouldn’t read is flagged and doesn’t hold up the rest. A fallback model on the same connection can pick it up, or you can read it on the Mac with “Read locally.”
What stops the whole pass
Only what is the same for every document: a rejected key, an empty balance, an answer in an unexpected format. The app tells you what happened instead of spinning forever.
FAQ
Questions about recognition
- Why five engines?
- Languages and hands differ. Tesseract knows languages Apple doesn’t; a cloud service may read a particular hand better; an AI model is worth a try when the others come up short.
- Do pages go out to the internet?
- Only to a cloud engine you switched on and confirmed yourself. By default, no.
- How accurate is it?
- It depends on the hand, the language and the page; we quote no number. Re-read a note with another engine and compare — the choice is always yours.
- What if an AI model can’t read a document?
- It gets flagged and the pass carries on. You can hand the document to a fallback model or read it on the Mac. Only a rejected key or an empty balance stops the whole pass.
- Is a document re-read on every search?
- No. What was read is stored and used by every feature. It’s read again only if the document changed or you picked another engine.