Local Secure Prompt Agent — AI skill for Claude Code

🛡 Privacy Agent

**A browser extension that redacts sensitive information from your prompts before they ever reach ChatGPT, Claude, Gemini, Copilot, or any other AI chat tool — entirely on your device.**

No cloud calls. No model downloads at runtime. No backend server. Detection and redaction happen synchronously, in your browser, the moment you paste something into an AI chat box.

---

Why

AI chat tools are increasingly where people paste logs, customer data, internal docs, and personal details to get help. Most of the time nobody notices that an email, a phone number, an API key, or a government ID slipped into the prompt along with everything else. Privacy Agent sits between your clipboard and the composer box and strips that out automatically, before it's ever sent.

How it works

Every paste into a supported AI chat site goes through two fully local, synchronous passes before it's allowed into the composer:

1. Pattern matching (regex) — catches structured, high-confidence PII: emails, Indian phone numbers, Aadhaar numbers, PAN numbers, IP addresses, API keys, generic secrets, and passwords.
2. NLP detection — catches names, locations, and organizations using `compromise`, a lightweight pure-JavaScript NLP library, combined with:
   • contextual phrase patterns ("my name is X", "I work at X", "I live in X")
   • a curated gazetteer of Indian states and major cities
   • a user-editable personal terms list — add your own name, employer, or city once, and it's always redacted regardless of phrasing

Both passes run before the paste event is allowed through, so the browser's default paste action is blocked and replaced with the sanitized version in one step. Nothing is sent anywhere to make this decision — there's no API call involved in detection at all.

What gets redacted

Category	Method	Example
Email addresses	Regex	[EMAIL]
Indian phone numbers