Mask the personal. Paste the rest.
Redact-a-bit swaps the personal details in a document for realistic fake ones, so the file still reads normally and the numbers still add up. It runs on your own computer.
Website · Local and offline · AGPLv3
You might want to share a bank statement with your accountant, paste a contract into an AI chatbot, or send records to a lawyer. Usually that means exposing your SSN, account numbers, and address right along with it.
Online redaction tools don't really solve it: you upload the very document you were trying to protect. Blunt [REDACTED] bars leave the file unreadable. Redact-a-bit swaps the sensitive parts for believable fakes instead, so the document still makes sense and an AI can still work with it. Your real information just isn't in the copy you share.
(The name is just redact a bit: take out what identifies you, leave everything else in place.)
Drop in a file, pick a mode (how each match is replaced) and a level (how much gets touched), then download a clean copy.
Name: Michael Thompson Name: Daniel Foster
SSN: 482-19-3756 → SSN: 837-44-1920
Account: 1029384756 Account: 6647201938
Balance: $48,210.55 Balance: $48,210.55 ← kept
That's Fake mode. It's not just replacement, though — the same SSN goes three ways depending on the mode you pick:
Fake 482-19-3756 → 837-44-1920 a realistic stand-in (best for AI)
Mask 482-19-3756 → XXX-XX-3756 last 4 kept
Label 482-19-3756 → [SSN REDACTED] a plain tag
- Levels. Light = the critical IDs (SSN, accounts, cards, passport/visa). Standard adds contact and identity (email, phone, address, DOB, named fields). Heavy adds amounts and ZIP codes, so a balance stays at Light/Standard but is replaced at Heavy.
- Detection. Built-in patterns catch structured IDs like SSNs, cards, and account numbers. In the desktop app, optional local GLiNER models also pick up names and addresses; they download the first time you use them.
- Repeatable. The same input always produces the same fake, so a document stays consistent with itself.
It handles PDFs and plain-text files (.txt, .md). There's no OCR yet, so scanned PDFs need a real text layer.
Start on the website. It always points to the current build.
Desktop app. Download for Windows is a single installer (around 50 MB) that runs the redactor and the optional detection models without needing Python. macOS and Linux are coming.
Python version (CLI / reference engine). For developers or headless/scripted use — same engine, no desktop install:
pip install gradio pymupdf # dependencies
python redactabit.py # opens the UI at http://localhost:7860
python redactabit.py statement.pdf -l 2 -m fake # or run it from the command lineYour document stays on your computer. It doesn't upload your file, ask you to sign in, or phone home. The interface only listens on 127.0.0.1, so nothing else on your network can reach it. The optional detection models download once, and after that it runs with the Wi-Fi off. It's open source (AGPLv3), so you can read the code and confirm all of this for yourself.
git clone https://github.com/prPMDev/redactabit.git
cd redactabit
# Python version
pip install gradio pymupdf && python redactabit.py
# Desktop app (needs Node, Rust, and your platform's build tools)
npm install && npm run tauri build # installers land in src-tauri/target/release/bundle/Tests run with pytest (the engine) and npm test (the desktop engine and detector).
Today. The redaction engine in Python and TypeScript, three levels and three modes, the regex floor plus GLiNER detection, PDF and text support, and repeatable fakes. It's US-focused for now.
Next. Signed installers for macOS and Linux, and bundling the model runtime so even the first run works offline.
Later. More countries (UK, India, the EU), keeping PDF layout intact, and OCR for scanned files.
AGPLv3 (AGPL-3.0-or-later). See LICENSE. Redact-a-bit is a passion project I don't intend to monetize, so copyleft is a feature, not a cost: use it, fork it, ship it, and keep your version open too. (True PDF redaction is built on MuPDF, which is AGPL.)
Built by prPMDev. The detection builds on GLiNER and the desktop app on Tauri. The site borrows its spirit from Handy.