guillaumemeyer/watermarks-remover

What it solves

This project provides tools to remove AI-generated provenance marks and watermarks from text and various file formats. It is designed for users who own their content and wish to maintain privacy or "hygiene" by stripping invisible identifiers used by AI vendors (such as Claude, Gemini/SynthID, and OpenAI) to track AI-generated content.

How it works

The system operates across three primary layers:

• Layer A (Deterministic): Uses Python scripts to remove invisible Unicode characters, exotic spaces, bidirectional overrides, and tag characters.
• Layer B (Statistical): Addresses token-sampling watermarks through agent-led rewrites or a specific rewrite_text.py hook that can use backends like Ollama.
• File Metadata: Strips C2PA, EXIF, XMP, and document properties from a wide array of formats, including images (PNG, JPEG, WebP, etc.), documents (PDF, DOCX, XLSX, etc.), and media files (MP4, WAV, etc.).

It can be deployed as a standalone HTTP service, a set of CLI scripts, or as an "agent skill" (a thin client) that allows AI agents (like Claude Code or Cursor) to trigger the cleaning process. It also supports "hooks" that can automatically scan and clean files written by an agent before they are saved to disk.

Who it’s for

• Content Creators: People who want to remove AI provenance markers from their own work.
• AI Agent Users: Users of Claude Code, Cursor, or Grok who want their agents to automatically clean the files they generate.
• Developers: Those needing a programmatic way to detect or strip AI watermarks via an HTTP API.

Highlights

• Multi-Vendor Support: Targets marks from Claude, Gemini/SynthID-Text, OpenAI, and open-LLM schemes (Kirchenbauer/Aaronson).
• Broad File Support: Handles everything from standard images and PDFs to Office documents and audio/video files.
• Agent Integration: Specialized installers for Claude Code, Cursor, and Grok, including a PostToolUse hook for automatic cleaning.
• Lightweight Core: The core service and scripts rely only on the Python 3.10+ standard library, with optional system tools (c2patool, exiftool, qpdf) for deeper cleaning.