PromptWard

PromptWard catches personal information in your prompts before it reaches ChatGPT, Claude, Gemini, Perplexity, or Mistral — and redacts it locally, on your device. No prompt text, no placeholder map, and no telemetry ever leaves your browser. HOW IT WORKS When you press send on a supported AI chat site, PromptWard briefly intercepts the prompt, runs a local ONNX model (Rampart) plus deterministic heuristics to find personally identifiable information — names, Social Security numbers, credit card numbers, email addresses, phone numbers — and shows you a side-by-side review: your original prompt next to the redacted version. Identifiers are replaced with reversible placeholder tokens like [PERSON_1] or [SSN_1]. The redacted version is the default. If you step away from the keyboard, it auto-sends after a 5-second countdown. You can also click to take manual control, send the original unchanged, or cancel. PRIVACY-FIRST BY DESIGN - Detection runs 100% on-device. There is no PromptWard server. Your prompts are never transmitted anywhere to be classified. - The placeholder map that would reverse [PERSON_1] back to the real name is held only in extension memory and cleared on navigation, tab close, or reload. - No accounts, no analytics, no telemetry. Settings store only your preferences, never prompt content. (Note: model responses are not yet rehydrated — if the assistant references a redacted value, you'll see the placeholder token rather than the original text. The placeholder maps that would enable this are kept per conversation and can be wired up later.) SUPPORTED SITES Works out of the box on ChatGPT (chatgpt.com, chat.openai.com), Claude (claude.ai), Gemini (gemini.google.com), Perplexity (perplexity.ai), and Mistral (chat.mistral.ai). Add any other site from the side panel's Custom Domains list — PromptWard will request permission for just that site. BY WHOM Built for anyone who has ever pasted a customer's details, a real SSN, or a colleague's contact info into a chat window and immediately wished they hadn't. Assistive, not a compliance guarantee: it can miss PII in unusual formatting and will occasionally flag harmless text. Always review the redacted version before sending. OPEN SOURCE MIT licensed. Source, model attribution, and the full privacy posture are at https://github.com/gduplessy/promptward.