Every step below happens inside your browser tab. Nothing is sent to a server, and no AI model is called — not even to detect names.
Step 1 Paste text or open a file
Paste an email, a contract, a support ticket, a bug report or a medical note into the editor. You can also drop a file onto the editor or use Upload File.
| Format | What happens |
|---|---|
.txt, .md, .csv, .log, .json | Read as plain UTF-8 text. Binary files are rejected. |
.pdf | Only the embedded text layer is extracted, page by page. A PDF without a text layer (a scan) is refused with an explanation, because OCR is not implemented. |
.docx | Plain text is extracted. Formatting is not preserved. |
.doc, .rtf, images | Rejected with a message asking for .docx or .txt. |
Step 2 Review what was detected
Detection runs locally, a quarter of a second after you stop typing. Each match appears in the findings list with a colour, a category and the placeholder it will receive; the preview on the right highlights the same spans inside your text.
Nothing is applied automatically. You can:
- uncheck an item to keep it untouched (for example a company name you deliberately want the AI to see);
- select any text in the editor and add it manually as a person, company or custom value;
- turn an entire category off in Detection settings if it produces noise for your document.
What is detected, and how
| Category | Method | Reliability |
|---|---|---|
| Email address | Structured pattern | High |
| IBAN | Country length + mod-97 checksum | High (invalid checksums are still offered, flagged “check this”) |
| Payment card | Length + Luhn checksum + brand prefix | High |
| Italian tax code / VAT | Pattern + control character; VAT requires explicit context | High |
| Phone number | Digit pattern with country, area or mobile prefixes | Medium — dates and IBAN digits are excluded |
| IP address | IPv4 octet validation, IPv6 structure | High; version numbers are excluded |
| URL | Scheme + host; parameters such as token, email, key are flagged | High |
| API key / secret | Provider prefixes, assignment context, JWT, PEM blocks, database URLs | High for recognised formats; medium for contextual assignments; masked in the preview |
| Date of birth | Date pattern plus surrounding context (“born”, “date of birth”, “nato il”) | Medium — a bare date is not flagged |
| Date | English, Italian, French and German month names, ISO and numeric formats, calendar validation | Medium — impossible dates and contract/version numbers are rejected |
| GPS coordinates | Decimal pairs, hemisphere notation and DMS; latitude ±90, longitude ±180 | High when labelled or hemisphere-marked; medium for bare decimal pairs |
| Person name | Local first-name and surname dictionaries, title prefixes, email local-part hints | Heuristic — review it |
| Company name | Legal suffixes (Srl, S.p.A., GmbH, Ltd, Inc…) | Heuristic — review it |
Step 3 Anonymize, then use the text
Press Anonymize Now (or Ctrl+Enter). The output panel contains the text with placeholders, and the mapping table below it shows which placeholder stands for which original value.
Placeholders are progressive and stable: the same value always gets the same identifier, so the assistant can still reason about the document. “John Smith wrote to John Smith’s manager” becomes “[PERSON_1] wrote to [PERSON_1]’s manager”.
Example
John Smith works at Northstar Solutions Ltd. His email address is john.smith@example.com. His IBAN is GB82WEST12345698765432. His phone number is +44 7700 900123. The delivery is on February 17, 2026, at 45.4642, 9.1900.
[PERSON_1] works at [COMPANY_1]. His email address is [EMAIL_1]. His IBAN is [IBAN_1]. His phone number is [PHONE_1]. The delivery is on [DATE_1], at [GPS_COORDINATES_1].
Restoring the original values
When the AI answers, paste the answer into Restore original data. The placeholders are swapped back using the mapping held in the tab’s memory. Matching tolerates extra spaces, different letter case and markdown emphasis such as **[PERSON_1]**.
Placeholders that are not in the mapping — for instance [PERSON_9], invented by the model — are left untouched and reported, never silently replaced with a guess.
The mapping disappears when the tab is closed or reloaded, or when you press Wipe all data. You can export it as a JSON file if you need to restore later; that file contains the original values, so store it accordingly.
Where this approach stops
- It is pseudonymization, not true anonymization: indirect identifiers can survive.
- Names and companies are heuristic. Unusual names, single first names and non-Latin scripts may be missed.
- Addresses, medical details and free-form identifiers are not detected automatically.
- No OCR, so scanned documents cannot be processed.
- No layout-preserving export: PDF and DOCX come out as plain text.
The interface repeats these limits where they matter. See the Disclaimer for the full statement.
Guides for specific tasks
The same workflow can be applied to a particular kind of material. These guides go into the details that matter for emails, documents, client work and AI prompts, including the values that still need a human decision.
- Anonymizing text before ChatGPT
- Anonymizing email threads
- Anonymizing documents
- Protecting sensitive data in AI prompts
Try it
The tool is on the home page, and the “Try an example” button fills the editor with a sample document so you can see the whole flow in a few seconds.