Skip to content

Repository files navigation

AIScrubber

AIScrubber is a browser-local privacy toolkit for cleaning sensitive text and images before they are shared with an AI service, issue tracker, or another person.

Open the web app · Report a bug

What it does

  • Text Scrubber replaces detected credentials, email addresses, IP addresses, payment cards, IDs, and other sensitive values with stable tokens.
  • Prompt Masker exports a local session key so placeholders in an AI response can be restored later.
  • Unicode Cleaner removes selected zero-width characters, Tag Plane tokens, unusual spaces, and confusables. Optional phrase rules clean common AI-style copy patterns.
  • Metadata Desk inspects JPEG, PNG, and WebP images and removes supported EXIF, GPS, PNG text, and C2PA-compatible metadata markers.
  • Media Redactor permanently burns blur, pixelation, or blackout regions into an exported image.

The browser tools process user content in local memory. The site fetches the repository's public GitHub star count, but no pasted text, files, mappings, or session keys are sent with that request.

Use the CLI

Run it without installing globally:

npx aiscrubber scrub ./incident.log --output ./incident.clean.log
npx aiscrubber mask ./prompt.txt --key ./session.aiscrub.json
npx aiscrubber unmask ./ai-response.txt --key ./session.aiscrub.json
npx aiscrubber clean-watermarks ./draft.md --output ./draft.clean.md
npx aiscrubber strip-metadata ./photo.jpg ./screenshot.png
npx aiscrubber inspect ./incident.log --json

strip-metadata currently writes sanitized copies for JPEG and PNG files. It does not rewrite PDFs or claim to verify C2PA signatures.

Use npx aiscrubber help <command> for command-specific examples.

Connect the MCP server

AIScrubber exposes scrub_text, mask_prompt, unmask_response, clean_ai_watermarks, and inspect_content over stdio.

{
  "mcpServers": {
    "aiscrubber": {
      "command": "npx",
      "args": ["-y", "aiscrubber", "mcp"]
    }
  }
}

Accuracy and safety boundaries

  • Detection is deterministic pattern matching, not semantic understanding. Review the output before sharing it.
  • Unicode findings do not prove which model created text or that a vendor watermark exists.
  • The Metadata Desk detects supported binary markers but does not verify cryptographic provenance signatures.
  • Keep .aiscrub.json session keys private. They contain the original values required for reconstruction.
  • Work on a copy of important files and verify exported images before deleting originals.

Develop locally

Requirements: Node.js 18 or newer.

npm ci
npm test
npm run build
npm run dev

Deploy the built static assets to Cloudflare Workers:

npm run build
npx wrangler deploy

License

MIT. See LICENSE.

About

Client-side AI privacy scrubber sanitize sensitive data before it reaches any LLM. No servers, no tracking, free forever.

Topics

Resources

Code of conduct

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Sponsor this project

Contributors

Languages