Guillaume Meyer's watermarks-remover Hits 15K Stars Stripping Claude's Invisible Marks
A new MIT-licensed agent skill and HTTP service scrubs Claude, Gemini, and OpenAI provenance signals from text, images, and documents.
Takeaways
- MIT-licensed watermarks-remover hit 15.7k GitHub stars stripping AI provenance signals from text and files.
- Three layers: deterministic Unicode scrub, statistical LLM rewrite, and C2PA/EXIF/XMP metadata cleaning.
- Covers Claude, Gemini/SynthID-Text, OpenAI and open-LLM Kirchenbauer-style watermarks.
- Ships as an agent skill plus stdlib Python HTTP service with OpenAPI spec and batch endpoints.
- Optional pixel-domain removal via CtrlRegen ControlNet regeneration for SynthID-class image marks.
- Rewrite pass is best-effort only; Anthropic has not released a public detector to verify results.
A GitHub project called watermarks-remover raced up the trending charts this week, collecting more than 15,000 stars within days of launch. Built by developer Guillaume Meyer, the tool ships an agent skill plus a Python HTTP service that strips provenance signals vendors have started baking into model outputs. Coverage from BleepingComputer notes the release landed days after Anthropic disclosed switching on invisible marks in everything Claude writes, and the project now targets multiple providers at once.
Three layers of scrubbing
The project splits the problem into three orthogonal jobs, which matters because different watermark classes need entirely different attacks:
- Layer A (deterministic text scrub): Python scripts remove invisible Unicode characters, exotic spaces, bidirectional control characters, and tag characters. Removal is 100 percent reliable.
- Layer B (statistical rewrite): An agent-driven rewrite pass targets token-sampling watermarks. This covers marks used by Claude, Google SynthID-Text, OpenAI attribution markers, and Kirchenbauer-type schemes commonly employed by open-source models.
- File metadata cleaners: C2PA manifests, EXIF, XMP and doc properties get stripped from PNG, JPEG, WebP, AVIF, HEIC, SVG, PDF, DOCX, XLSX, PPTX, EPUB, HTML, Markdown, MP4, WAV, MP3 and more.