Work with AI. Keep client data on your desk.
Redact for AI finds names, identifiers and addresses in documents and replaces them with consistent placeholders — on your computer, with no network. When the AI answers, one click restores the real names.
▲Free, always- Person_1
- Client_1
- SSN_1
- Person_2
- Client_2
▲The software works with the connection off. Nothing is uploaded, ever.
Four steps, one review.
Open the document
PDF, Word, text or scans — OCR runs locally.
Anonymize
Every detected item becomes a placeholder: Person_1, Client_1.
Review and send
Check the highlights, fix what's missed, copy the anonymized text into your AI.
Restore
Paste the AI's draft back: real names return in one click.
Only the anonymized text goes to the AI.
The original never leaves its folder.
Review every document.
Automatic detection is never 100%. The preview shows exactly what will be replaced.
Never share the _private folder.
It holds the maps that link placeholders back to real names.
100% local
No cloud, no server, no account. Verified: it runs with the network cable unplugged.
Review with highlights
See every replacement before anything leaves the app. Add what's missed with one selection.
Roles for the AI
Mark who is the Client and who is the Opposing party — placeholders become readable context.
Reversible
The AI's draft comes back with real names restored, locally.
Four legal systems
US identifiers (SSN, EIN), Italian (codice fiscale, P.IVA), Brazilian (CPF, CNPJ, OAB), Spanish (DNI, NIE, CIF).
Scans included
Built-in OCR for scanned PDFs, in English, Italian, Portuguese and Spanish.
Interface and detection engines for four jurisdictions.
Five pages: install, review, restore. Pick your language.
Free, always.
Redact for AI is free and will stay free. Leave your details to download — we use them to tell you about updates and other Living Off AI tools.
▲First launch requires one extra approval on macOS and Windows — instructions appear after the download starts.
macOS 13+ or Windows 10+. First start loads the language models (30–60 seconds).