Format-only detection without a checksum would produce too many false positives, so each detector validates before redacting.

What you (the model) provide

• names — personal names you've identified by reading the text. The tool does no NER. Pass full names like "Jean Dupont", not just "Jean", to avoid false matches against common words.
• custom_terms — anything else you want gone: company names, addresses, project codenames, etc.

There's also a per-chat config field Always-redact terms (string array): things that should be redacted in every call regardless of model input — e.g. your own name, your home address, an employer name. Useful as a safety net.

Stable pseudonyms

Same value → same pseudonym, throughout one call:

(The model is expected to pass names: ["Jean", "Jean Dupont"] — but [NOM_1] is reused because the value matched in the text is the same once dedup runs.)

The response includes a mapping so you can build a "key" file alongside the redacted version if needed:

Set include_mapping: false to omit it (e.g. if you're going to forward the redacted text somewhere and don't want the mapping in your context).

Composing with the filesystem plugin

The intended workflow:

• read_file({path: "contract.md"}) — get the original.
• anonymize_text({text, names: [...]}) — model identifies names, plugin redacts.
• write_file({path: "contract.redacted.md", content: anonymized}) — save the clean version.

Configuration

Quality

The detectors are checksum-validated, not regex-only. This is a deliberate trade-off: high precision at the cost of some recall on synthetic or poorly-formatted data.

Measured on the FR subset of AI4Privacy pii-masking-400k (~40 000 French-locale samples, local eval only — see eval/run-ai4privacy.ts):

Take-away: this plugin is the right tool when you want zero false redactions of look-alike data (reference numbers, internal IDs). It is not the right tool when you need to scrub arbitrary digit sequences — there's no validation signal there, by design.

Run the suite locally:

Install

One-click (macOS, Windows): Run in LM Studio

Linux (AppImage):

After install: in any chat, click the tools button and enable anonymize.

For contributors

See docs/ for design notes, per-detector algorithms, evaluation methodology, dataset licensing, and the roadmap.

License

MIT