Problem
People paste everything into AI chatbots: emails to rewrite, contracts to summarize, logs to debug. And “everything” includes names, phone numbers, addresses and IDs that never needed to be part of the prompt. Nobody decides to leak personal data; they decide to make an email sound friendlier, and the leak comes free.
How it works
- A Chrome Manifest V3 extension intercepts paste events on ten AI chat platforms (ChatGPT, Claude, Gemini and friends) and scans the clipboard text before the page ever sees it.
- Detection runs a Hugging Face token-classification model in the browser via transformers.js and ONNX Runtime, accelerated with WebGPU (WASM fallback included) and hosted inside an offscreen document so the heavy work never touches the page.
- Eight PII entity types are detected and masked inline; you see exactly what was redacted before the text goes anywhere.
- Nothing leaves the machine. The model is the privacy boundary: no API calls, no server, no telemetry.
Design decisions
- Fail-open. If the model isn’t loaded or inference errors out, the paste goes through untouched. A privacy tool that randomly eats your clipboard gets uninstalled within the hour, and an uninstalled privacy tool protects no one.
- Client-side or nothing. Sending your text to a cloud API to check it for private information is a punchline, not an architecture.
- Shipped like a product. Packaged for Chrome Web Store release, not left as a repo whose README opens with “clone and pray”.