Claude's own watermark cannot be removed by any tool — Anthropic states plainly that nothing is added to the text and there are no hidden characters. What you can remove is the other layer: invisible characters and file metadata. Why →
Nothing cleaned yet
Paste text on the left and press Clean the file. What was removed shows up here, along with a count.
For images this panel also lists what was kept — the colour profile, for instance, so cleaning never changes how a picture looks.
Read this first. Claude's own text watermark cannot be removed by any tool — including this one. It is a statistical pattern in how Claude chose its words, not characters stored in your file. Anthropic says so plainly: "Nothing is added to the text and there are no hidden characters." There is nothing in the text to delete.
What this tool does remove is the other layer — the things that really are bytes in your file: invisible zero-width characters, bidirectional control marks, hidden class attributes, and image metadata (EXIF / XMP / C2PA). Those have many sources, and only some of them are AI-related.
If you came here for a button that strips Claude's statistical watermark, that button does not exist — anywhere. Most sites selling a "Claude watermark remover" let you believe it does. Full explanation →
This is a cleaner for files you already have. It reads the file you hand it, strips the machine-readable layers inside it, and gives it back — all inside your browser. It has no way to reach anything else: there is no account to sign into, no server that could receive a request, and no share link it could open on your behalf.
That boundary is deliberate. Removing the provenance record from your own export is ordinary editing — no different from retyping a paragraph. Using a tool to strip someone else's attribution and pass their work off as your own is a different thing, and this tool is not built for it: it cannot fetch a file it was not given, and it does not claim to make anything untraceable.
No — and neither can any other tool. Claude's watermark is a statistical pattern in how the model chose its words, not characters stored in your file. Anthropic's own documentation says "nothing is added to the text and there are no hidden characters", so there is nothing in the text for a cleaner to delete. Full explanation →
The layer that really is bytes in your file: invisible and zero-width Unicode characters, bidirectional control marks, hidden HTML class attributes, and image metadata — EXIF, XMP, and the C2PA credential Claude attaches to the images it produces. Those have many sources, and only some of them are AI-related.
For the layer this tool touches, treat it as ordinary editing: zero-width characters and file metadata are not copyright notices or licence terms, and removing them from your own file is no different from retyping a paragraph. The statistical watermark is a separate question, and Anthropic's documentation describes no prohibition on editing text the model produced. In practice the constraint is rarely a statute — it is your school's academic-integrity policy, your employer's disclosure rule, or a journal's submission terms, and no cleaner changes what those say. Stripping metadata from someone else's file to pass it off as your own is a different thing entirely. This is not legal advice; if the stakes are real, ask someone qualified.
Anthropic says it does not: "Nothing is added to the text and there are no hidden characters." Hidden characters do turn up in real documents, but they come from copying through web pages and chat interfaces, PDF text extraction, autocorrect, and pasting between apps — not from the model inserting them. Find them in your own files →
Not in the same way — and only one provider has published a text watermark at all. Anthropic's is statistical: a bias in how the model picks each next word. OpenAI's provenance documentation describes C2PA content credentials attached to files and images, not a text watermark; the invisible-character stories that circulated around ChatGPT in 2025 were character-layer artifacts picked up in the interface, not something the provider embedded. The distinction decides whether a cleaner can do anything: metadata sitting on a file can be deleted, a bias in word choice cannot. The two layers →
No, and it is worth being blunt about that. Stripping invisible characters does not change how the text was written, and detectors do not work by hunting for zero-width spaces. If a site promises that cleaning characters will beat a detector, it is selling something it cannot deliver. The longer answers →
No. Everything runs in your browser — there is no backend that could receive your file. You can confirm that in about fifteen seconds: open DevTools, switch to the Network tab, clean a file, and count the requests. How to check →