Watermarks-remover is a Python tool and agent skill that strips AI provenance marks from text and files across multiple vendors and formats.
The tool addresses the problem of removing invisible markers that AI systems embed in generated content for tracking and authentication purposes. It works across three layers: Layer A uses deterministic Python scripts to remove invisible Unicode characters, exotic spaces, bidirectional text, and tag characters; Layer B handles statistical text watermarks through agent-based rewriting and optional hooks; and a file layer removes C2PA manifests, EXIF, XMP, and document properties from PNG, JPEG, SVG, PDF, DOCX, ODT, HTML, and Markdown files. It targets provenance marks from Claude, Gemini with SynthID-Text, OpenAI, and open-source LLM implementations using Kirchenbauer-style watermarking.
The tool suits developers and content creators who own generated text or images and want to remove vendor-specific tracking marks for privacy or content hygiene. It works as both a standalone Python script suite requiring only Python 3.10 and standard library, and as an agent skill invoked through natural language. The project integrates optionally with external tools like c2patool for C2PA inspection and exiftool for residual metadata removal. For image watermark detection, it can optionally load an external SynthID scorer from a separate checkout, though pixel-level watermark removal remains out of scope. The tool explicitly does not address training backdoors or pixel-domain watermark removal for most vendors.
The project maintains focused scope boundaries with clear documentation of what each vendor's marks look like and which removal strategies apply. Development activity shows deliberate architectural choices around optional dependencies and external integrations, with careful licensing considerations for bundled versus runtime-loaded components. The codebase distinguishes between detection and removal capabilities, documenting which are in scope and which are not.
