watermarks-remover

watermarks-remover is an open source, privacy-first toolkit for detecting and stripping AI provenance marks from content the user owns. The repository focuses on three categories: invisible Unicode and unusual spacing marks in text, statistical text watermark patterns that may require rewriting, and metadata or provenance containers such as C2PA, EXIF, XMP, document properties, and media metadata. It is a security and content-hygiene tool rather than a creative generator. The project ships a local service, an audit library, deterministic cleaners, an optional rewrite hook, an agent skill installer, and integrations for environments such as Claude Code, Cursor, Cowork, claude.ai cloud sessions, and routines. The README says the agent skill itself is markdown and calls the local service over HTTP. Users install it with commands such as `python3 install_skill.py --skill remove-ai-marks --target claude-code`, then invoke the skill from their AI coding or writing environment when they want to check or clean files. Coverage is broad. For text, the tool audits invisible Unicode, exotic spaces, bidi controls, tag characters, and other hidden marks. For files, it targets common provenance and metadata containers across PNG, JPEG, WebP, AVIF, HEIC, BMP, GIF, TIFF, SVG, PDF, DOCX, XLSX, PPTX, EPUB, ODT, HTML, Markdown, MP4, MOV, M4A, M4V, WAV, MP3, and FLAC. The hook can run in check mode, which reports findings without changing files, or clean mode, which strips deterministic marks in place and tells the model that the file changed on disk. The tool is useful for developers, creators, and security teams who need to inspect their own generated assets before sharing them, publishing them, or passing them through systems that react poorly to hidden metadata. It can also help teams understand what provenance data is attached to generated files and whether their release process leaks document properties or hidden text markers. The README is clear about one limitation: no hook can rewrite an assistant’s chat message after it has already been sent. Cleaning applies to files and supported content paths. Because watermark removal can be misused, the safest framing is ownership and audit. watermarks-remover should be used on content you own or are authorized to process, and teams should preserve required disclosure or attribution obligations where policy, contract, or law requires them. The repository is MIT-licensed, implemented primarily in Python, and had more than nineteen thousand GitHub stars when reviewed, which makes it a notable entry for builders tracking provenance, C2PA, and AI content hygiene tools.