Next-Gen Browser-Local Privacy Desk, Developer CLI & Model Context Protocol (MCP) Server
🌐 Live Web Suite · 💻 CLI Reference · 🤖 MCP Server Setup · ✨ 5 Privacy Engines · 👨💻 Founder
AIScrubber is an offline-capable, browser-local privacy desk and developer toolkit engineered for the exact moment before sensitive data travels to AI frontier models (ChatGPT, Claude, Gemini), public GitHub issues, or email threads.
All user-provided content is processed in local memory and is never uploaded to AIScrubber. The website uses privacy-respecting page analytics and fetches the public GitHub star count; neither request contains text, prompts, files, or redaction data.
- Scans input text and incident logs against 9 built-in detector classes: email, phone, validated IPv4/IPv6, URLs, Luhn-valid payment cards, provider credentials, system IDs, US SSNs, and Verhoeff-valid Aadhaar/PAN formats.
- Recognizes OpenAI, GitHub, Stripe, Google, GitLab, SendGrid, npm, Slack, AWS, JWT, bearer-token, and private-key credential shapes without matching split-line fragments.
- Supports Custom Keyword & Regex Rules drawer.
- Interactive side-by-side Diff Inspector with token tooltips.
- 1-Click Dictionary Key Export for reversible token tracking.
- Step 1 (Mask & Structure): Detects credentials and endpoints in raw prompts and replaces them with semantic constants (
{{API_SECRET_1}},{{DATABASE_URL_1}}). - Step 2 (Key Export): Exports a downloadable
.aiscrub.jsonsession key. - Step 3 (Query AI): Paste the masked prompt to ChatGPT or Claude.
- Step 4 (1-Click Reconstruct): Paste the AI's generated response and session key to unmask all original variables back into working code.
- Strips invisible Unicode zero-width watermarks (
\u200B,\u200C,\u200D,\uFEFF,\u2060) embedded in Anthropic Claude & LLM outputs. - Normalizes non-standard synthetic spaces (En, Em, Thin, Hair, Narrow No-Break, Fullwidth spaces).
- Reverts confusable Cyrillic/mathematical homoglyphs to Latin standards.
- 4 Modes: Aggressive (All Tiers), Claude & LLM Output Clean, Code Safe (Preserves Indentation), and Invisible Unicode Only.
- In-browser parsing of JPEG (EXIF/GPS/IPTC/XMP), PNG chunks (
tEXt/caPI), PDF/Infodictionaries, and Audio ID3 tags. - C2PA Content Credentials [CR] Inspector & Signer: Detects cryptographic provenance manifests from ChatGPT (DALL·E 3), Nano Banana, Adobe Firefly, and Google Imagen, and allows custom binary tag signing.
- 1-Click Stripper: Re-encodes clean pixel buffers and wipes C2PA tracking fingerprints.
- HTML5 Canvas interactive tool to draw redaction boxes on screenshots and images.
- Tools: Gaussian Blur, Pixelate (Mosaic), and Solid Blackout.
- Undo/redo history stack and clean high-resolution PNG export.
Run the standalone CLI with zero installation via npx:
# Strip Claude & AI invisible Unicode zero-width watermarks
npx aiscrubber clean-watermarks ./claude-output.md -o ./clean-article.md
# Scrub sensitive incident logs into safe numbered labels
npx aiscrubber scrub ./logs/production-crash.log --output ./logs/clean.log
# Mask prompt secrets for AI and save session key
npx aiscrubber mask "Connect postgres://admin:P@ssw0rd@db.internal:5432" --key session.aiscrub.json
# Unmask AI response with your session key
npx aiscrubber unmask ./ai-response.py --key session.aiscrub.json --output ./final-code.py
# Inspect file or text for hidden EXIF, C2PA, or leaked API keys
npx aiscrubber inspect ./production-dump.log
# Strip metadata & C2PA manifests from documents and images
npx aiscrubber strip-metadata ./photos/*.jpg ./reports/*.pdfGet instant interactive usage instructions for any command:
npx aiscrubber help clean-watermarks
npx aiscrubber help scrub
npx aiscrubber help mask
npx aiscrubber help unmask
npx aiscrubber help strip-metadata
npx aiscrubber help inspectConnect AIScrubber directly to Claude Desktop, Claude Code CLI, or Cursor as an official Model Context Protocol (MCP) server.
Add the following to your claude_desktop_config.json:
{
"mcpServers": {
"aiscrubber": {
"command": "npx",
"args": ["-y", "aiscrubber", "mcp"]
}
}
}clean_ai_watermarks: Strips invisible Unicode zero-width watermarks (Anthropic Claude & ChatGPT markers), normalizes non-standard spaces, and reverts homoglyphs.scrub_text: Sanitizes raw logs and text into[EMAIL_1],[SECRET_1],[IP_1].mask_prompt: Masks secrets into{{KEY_1}}constants and outputs session key JSON.unmask_response: Restores original variables back into returned AI output.inspect_content: Scans text or files for leaked credentials, API tokens, invisible watermarks, and C2PA tracking manifests.
| Characteristic | Specification |
|---|---|
| Execution Environment | Client-Side Browser Memory / Local Node.js Process |
| Sensitive Content Uploads | 0 — processing stays in browser memory or the local Node.js process |
| Non-content Requests | Page analytics and the public GitHub star count; no pasted text, files, mappings, or keys |
| Storage | Ephemeral RAM only (LocalStorage used only for dark/light UI preference) |
| License | Open Source MIT License |
Designed and built by Poorvith M P, 19-year-old student developer & founder based in Bengaluru, India.
- Portfolio: poorvithmp.com
- GitHub: @poorvith-mp
- X (Twitter): @poorvithmp
- LinkedIn: linkedin.com/in/poorvithmp
Distributed under the MIT License. See LICENSE for details.