Runs on this device. Nothing is uploaded.

Redact a file before you share it.

Drop in a spreadsheet or a text file. People, contact details, ID numbers and your organisation’s names are found and replaced in your browser, by the same rules the ProcessScout platform uses. The page shows you that nothing was sent.

Beta: it misses some. Always review the output before you share it. The receipt lists every change it made, so you can check. How accurate is it?

Your organisation

The names that would identify your company in this file. They stay in this tab and are cleared when you close it.

Add names to hide

One per line.

Company name Include former names and abbreviations. Subsidiaries and trading names Email and web domains Addresses at these domains become client.example. Sites, offices and plants Customers Suppliers and partners Products and brands Project and code names Job titles only one person holds For example, the only Chief Refinery Officer. Other words to hide Deal names, client references, anything else. Words to keep visible System names or terms the rules would otherwise hide.

How much to hide

Level StandardPeople become pseudonyms (“Person 3”). Emails, phone numbers, ID and bank numbers, dates of birth and health details are removed. Your organisation, customers, suppliers, sites and code names become aliases. StrongEverything in Standard, and salaries, amounts, prices and margins become ranges (“£40–60k”).

Names model OnAlso finds people, companies and places that patterns miss. A 124 MB download, kept in your browser’s cache. OffPatterns and name lists only. Faster, and the default on phones.

Your file

Getting the redaction engine ready…

The file picker opens when the page has locked itself.

Choose a file

CSV, TSV, TXT, Markdown, JSON or Excel (.xlsx), up to 50 MB. Word, PDF and email are coming.

Beta · measured, not promised

How accurate is it?

We test the engine on 300 files we never tuned it on: 13 formats from 12 fictional companies, with 9,016 personal identifiers planted in them. Today it hides about 9 in 10 of those identifiers. We publish these numbers and they improve with each release.

89–92%of personal identifiers hidden, with the platform’s names model

81%in Quick mode, with no names model

100%of emails, UK NI and NHS numbers, passports, IBANs and sort codes

88%of organisation names, sites and domains, when you list them in step 1

Held-out benchmark v0, 19 October 2026, ProcessScout engine with spaCy names recognition except where marked Quick mode. Ranges are the spread across settings.

Names are the weakest part: about 8 in 10 are found. Given names that are also words (“Will”, “Joy”), “Surname, Given”, “Dr Dunstan” and “A. Malik” are the common misses.
More than half of those test files still had at least one name, phone number or ID left in clear. That is why the receipt lists every change, and why you should read the output before sharing it.
Site codes in asset tags, a site’s town on its own and job titles only one person holds are often left visible. Add them in step 1, under “Other words to hide” and “Job titles only one person holds”.
The browser’s names model (bert-base-NER) hasn’t been scored on this set yet. On a smaller set of names it missed fewer than the platform’s model. We’ll publish its numbers next.

Check it yourself

Switch your network off. Once the page says it is locked, turn off wifi and drop in a file. It still works, because nothing it does needs a server.
Watch the traffic. Open your browser’s developer tools on the Network tab before you drop a file. Dropping it adds no requests; the counter on this page says the same.
Read the lock. The page adds a Content-Security-Policy of default-src 'none' after loading. Your file is read inside workers started after that, which inherit it. You can see the policy in the page’s elements.
Compare the fingerprints. Every file the page loaded is listed with its SHA-256 and was checked against the build manifest before use.

What it doesn’t do yet

Word, PDF and email files are coming next. PowerPoint, images, scanned documents and handwriting aren’t read.
Detection is probabilistic. Names in unusual forms, handwriting and faces may be missed. The receipt lists every place that was changed and the page checks the output for anything it hid elsewhere. Always review before sharing.
The browser finds names with a different model from the platform’s. We publish the comparison for each version.
Pseudonyms are consistent within this tab only. Nothing is remembered between visits.
In Excel files, formulas that contain text or point at other sheets are replaced by their values. Comments, document properties, charts, pivot tables and links are removed and counted in the receipt.

The same protection runs inside ProcessScout.

Every source your team adds is redacted to your level before anything is stored or read by a model.

See how it works Free tools