Nothing is added to your writing. The word choices themselves are the mark. Here is what it detects, what it cannot, and the honest version of removing it.
Claude now leaves an invisible watermark in the text it writes. Nothing is added to your writing. No hidden characters, no symbols to strip out. The word choices themselves are the mark.
Most people came here for the tool, so it is first. The part actually worth understanding is right after it.
Open source, MIT licensed. It fully removes invisible characters and file metadata, and that half is lossless. It cannot promise the word-choice watermark is gone, and the project's own README says so. Runs as an agent skill, or locally with python3 service/scripts/server.py --host 127.0.0.1 --port 8765 and no dependencies.
Three layers, and the difference between them is the whole story.
Lossless
Invisible Characters
Zero-width spaces, bidirectional marks, exotic spaces, tag characters. Your visible words do not change at all. Genuinely useful for anyone who has pasted text and had strange artifacts ride along with it.
Lossless
File Metadata
Strips C2PA content credentials, EXIF, XMP and document properties from PNG, JPEG, PDF, DOCX, SVG, HTML and Markdown. The content itself is untouched.
Best effort only
Statistical Rewriting
This is the layer aimed at the actual watermark, and the one to be clear-eyed about. The only way to affect the pattern is to change the words, so the README states plainly that removal means rewording, that rewording degrades the copy, and that no tool can honestly certify a vendor's detector will come back clean.
The project says it is for content you own. Worth adding: the EU obligation falls on the AI providers, not on you as a writer. Passing off AI work as your own where you agreed not to is the same problem it always was, tool or no tool.
A model writes one word at a time. Take "the weather today was cold and." The next word could be overcast or grey. Both work, and the meaning does not change. Normally that coin flip is settled by a plain random number.
Watermarking does not change the available words. It changes the source of the randomness. A secret key plus the few words before it decide which option gets picked. The output still looks random to you, but a key holder can read back through a long passage and check whether those choices match the pattern the key would produce.
So nothing is inserted, there is nothing to find by looking, and nothing to delete. It carries no information about you, your account or your company. It costs no extra tokens, and in side-by-side testing readers could not tell watermarked output from unwatermarked.
This is not a Claude-only thing either. As of August 2nd, EU rules require providers serving that market to mark AI content, and around 190 companies signed the same Code of Practice. Anthropic applies it globally because there is no clean way to scope it by region yet.
This is the part worth knowing before someone waves a result at you.
It can
Suggest Claude Was Involved
Given enough text, a key holder can assign a probability that Claude touched it. Longer passage, higher confidence.
It cannot
Tell Writing From Editing
It cannot separate "Claude wrote this essay" from "I wrote this and Claude fixed my commas." Both can come back flagged, and that single limitation is the entire argument.
It cannot
Work On Short Text, Facts, Or Code
Fewer word choices means less signal, so a caption is basically nothing. And where only one right answer exists there is no free choice to hide a pattern in. After "Newton's most famous work was called Principia" the next word has to be "Mathematica."
Anthropic has said a detection API is coming but has not shipped one, and has published no accuracy thresholds or dispute process. If a school or an employer starts treating a result as proof, that is a policy problem, not a technology one.
This prompt does the half of the job that is genuinely safe. It cleans what can be cleaned losslessly, tells you the truth about what it cannot see, and makes you answer three questions before you degrade your own writing chasing a pattern that may not even be there.
Marking machine-generated content is reasonable. The gap is that a mark proves involvement, not authorship, and there is no appeals process yet. The most useful move is keeping more of your own thinking in the work. Less to strip, and better writing anyway.