Stripii

Stripii takes the personal details out of a document before you paste it into an AI tool: names, dates, addresses, account and ID numbers. The medicine and the money stay, because a lab report with its results removed is not worth asking about.

It came out of the healthcare work, where the problem is obvious and constant. Someone uploads a medical record or a bank statement to get help reading it, and hands over everything else in the document at the same time.

What it does

There is a free Mac app: drop in a PDF, a scan or a phone photo, and get back text to paste and a redacted copy to save. There is also a command line tool built on the same engine, open source under MIT, with profiles for the HIPAA Safe Harbor identifier list and for financial statements.

Redaction is painted into the page rather than drawn over it, so nothing under a box can be copied out. Then Stripii reads its own output back with on-device OCR and looks for every detail it removed. If something can still be read, the result is held until you look.

Why it is an app and not a website

A website would need your document on a server, which is the entire thing this is trying to avoid.

The next option was a local web interface, and that failed for a reason worth recording: WebKit will not render inside a sandbox with no network access. The offline promise chose the technology. The app has no network entitlement, so the operating system refuses any connection it tries to make.

What the audit found

Before release I had the core audited. It returned 14 findings and all of them are fixed. One is worth stating plainly, because it is the kind of thing a privacy tool must not do.

Images sitting inside ordinary text-layer PDFs were not being redacted at all, and the document was still graded ready to share.

The read-back exists because of that finding. Those pages are now OCRed and checked again, and a page that still shows something is held rather than released.

Where it falls short

Recall on my synthetic benchmark is 0.927. Not 1.0, and I am not going to round it up.

Names in running prose without a label can slip through, and so can a description that identifies somebody without naming them. This is a de-identification tool. The health profile follows the Safe Harbor list, which is a method and not a certification, and nothing here makes anything HIPAA compliant.

It needs macOS 14 or later on Apple silicon. The app is not notarised, and its source is closed for now. The engine and the command line tool are open.

The download, the install steps and the full list of limits are on the Stripii site.