Claude Watermark Remover & Detector

A Claude watermark remover strips invisible Unicode characters out of text so what you paste is exactly what you see. Invisible characters are real codepoints that occupy no width on screen: zero-width joiners, the 256 variation selectors, bidirectional controls, tag characters, and private-use codepoints. They survive copy and paste, they travel through email and CMS fields, and they are counted by every program that measures a string, which is why a JSON key stops matching, a password comparison fails, or a diff shows two identical-looking lines as different.

This tool does two jobs. The detector scans your text and reports every invisible character it finds, grouped by class, with its exact codepoint and how many times it occurs, so you can see what is in there before you change anything. The remover then strips whichever classes you choose and hands back clean text. Both run entirely in your browser - nothing is uploaded, logged, or stored, which matters when the text you are cleaning is an unpublished draft, a contract, or a customer record.

One thing worth being precise about, because most pages on this topic are not. Claude's official text watermark, introduced in August 2026, is not made of invisible characters. It is a statistical watermark: token choices are nudged according to a key Anthropic holds, and the signal is spread across hundreds of words rather than stored in any single character. That means no client-side tool can remove it by deleting characters, and no client-side checker can confirm its presence or absence without that key. Anyone claiming otherwise is describing something else.

What this tool removes is the invisible Unicode that genuinely does appear in AI output, in text pasted from documents and PDFs, and in anything that has passed through a rich-text editor. That is a real and separate problem, and it is the one that breaks your code. The same scan works on output from Claude, ChatGPT, Gemini, and any other model, because invisible characters are a property of the text, not of the model that produced it.

Use it to clean your own drafts before they go into a database, to sanitise text before a string comparison, to check why a JSON payload will not parse, or to audit content you have been handed and cannot vouch for. The findings table doubles as an explanation: each class says what those characters do and why their presence matters.