Mask

Work text is full of client names, account numbers and addresses, and it gets pasted into AI tools every day. Mask swaps those details for placeholders in your browser before anything leaves your machine, and puts the real values back into the reply. The mapping never leaves your device. Masking lowers exposure - it is not anonymisation. Tested on 12 synthetic business texts: 26 of 26 seeded values (100%) caught. Happy to answer questions.

The unmask step is what makes this usable day to day: hiding a client's name is easy, but getting the AI's reply back with the real names in the right places is where manual redaction falls apart. Do placeholders stay consistent within a session, so the same client is always the same token when I ask a follow-up?

26 of 26 caught is reassuring. How does it handle names that are also ordinary words, like a client called "Apple" or a person named "Will"? And do custom term lists stay only on this machine?

Keeping the placeholder mapping on the device is a useful boundary. Is there a review step that highlights uncertain matches before copying or sending the masked text? For customer emails, being able to catch a missed name or correct an over-redacted sentence before the relay runs would be particularly helpful.

26/26 on synthetic texts is a nice start but I'd want to see it on messy real emails, signatures and forwarded chains are where names usually slip through.

Doing the masking in the browser so the original text never leaves the machine is a sensible design, and the Ask-the-AI relay that swaps real values back into the reply makes it practical for daily use. The custom term lists for client names and codenames are a nice touch. How does it handle ambiguous cases, like a person's name that is also a common word, and can the term lists be shared across a team?

Doing the masking client-side and keeping the placeholder mapping on-device is the right call for this use case. How does it handle names that are also common words (a client called "Apple" or a person named "Hope"), and is there a quick way to review and override what gets masked before copying?

The second check on the relay that nothing raw slipped through is a nice touch. When the detector misses something, like a client's nickname or a misspelled surname, can I add it to my term list straight from the masked preview so it gets caught next time? A quick count of what was masked by type before I copy would also be a handy sanity check.

Masking pasted text makes sense, but in my experience a lot of client data goes into ChatGPT as attachments: PDFs, exported email threads, spreadsheets. Does Mask handle files too, or only text in the box? And when it inserts into ChatGPT or Claude, does it type into their editor directly or go through the clipboard?