How it works, in detail
Everything below happens on your Mac. The text and files you check are never uploaded, and checking makes no API calls. The app goes online only to activate, and when you choose to send feedback, after showing you exactly what will be sent.
1. Exact patterns go first
ID numbers, card and bank numbers, emails, phone numbers and IP addresses are found by exact patterns, as are labelled dates of birth and passport, licence and account numbers. The model can never overrule a pattern's hit.
2. A local model reads for meaning
A 4-billion-parameter language model, running on your Mac, judges what patterns can't see: a first name in passing, or a symptom in a note. It leans towards flagging, because a false alarm costs a second look and a miss costs a leak.
3. Clean-looking text gets a second look
The extra check for names and health, on by default, asks two more questions of anything the model clears: does it name a specific person, and does it describe someone's health?
4. Real documents, part by part
Paste text, or import a .txt, .csv or .pdf. Long documents are checked part by part, with the parts overlapping, so nothing near a break or past page one slips through. Tables such as CSV exports are listed row by row, and PDF pages with no text are reported as not checked.
5. One clear answer
Found, none found, or not sure, and how sensitive: High for ID numbers, card and bank details, health details and dates of birth; Low for names, emails and phone numbers. You get a list of what was found and where, marked in your text. A part it couldn't check is never reported as clean.
6. Realistic fakes, then a second check
Each detail found can be swapped for a realistic fake, taken from ranges set aside for examples, such as 555-01xx phone numbers and example.com emails, so the fakes can't belong to a real person and the text still reads naturally in an AI tool. One person stays one person throughout, a .txt or .csv keeps its layout, and you can copy or download the result.
Then check the result again; the check looks past the app's own fakes. If anything is still flagged, look at it before you share. Anything that wasn't found stays in the text, and the app says so plainly.
What it can't do
Detection is probabilistic. It can miss personal information, especially names in ordinary sentences, and it can flag things that aren't personal. Its results don't show that a document may lawfully be shared under any privacy law. Review the output yourself before you share it, and don't rely on the app as your only safeguard.
Want the test results, with numbers? Ask us.
Want to try it?
Early-access invites go first to the first 100 people on the waitlist, and to anyone who brings three others.
No spam, and we never sell or share the list: your email is used only to contact you about access. We count visits without cookies.