Claude Metadata Checker & Remover Tool

It does two things. The checker opens a file — or a block of text — and reports every provenance marker it can point at in the bytes: C2PA manifests, EXIF blocks, XMP packets, document properties and invisible Unicode characters. The remover then strips those markers and hands the file back. It supports PNG, JPEG, WebP, SVG, PDF, DOCX, XLSX and PPTX, and the remover button only appears once the checker has actually found something to remove.

Anthropic marks Claude's output two different ways under its EU AI Act Article 50(2) transparency commitment, which applies to new models from 2 August 2026. Generated files such as .png, .jpg and .svg carry signed provenance metadata following the C2PA open standard. Text carries an imperceptible watermark woven into the wording itself, which survives copy and paste but degrades under heavy editing or translation. This tool reads and strips the first kind.

No — and neither can any other tool, whatever it claims. Anthropic has not published the watermarking algorithm or a public detection endpoint, so nothing outside Anthropic can verify it. Products advertising that ability are guessing from writing style, and style-based detectors misfire hardest on non-native speakers and on technical writing, which is exactly the wrong failure mode for academic or compliance use. What this tool does on the text side is exact rather than probabilistic: it finds invisible Unicode characters — zero-width spaces, bidi controls, tag characters, variation selectors — and decodes any data hidden inside them. A character is either present in the string or it is not.

No, because nothing is re-encoded. Each format has its own parser that rewrites the container and copies the payload through byte for byte, so a cleaned PNG's compressed image data is bit-identical to the original and there is no generational quality loss. Anything that affects rendering is kept deliberately: ICC colour profiles, JFIF density, PNG transparency and gamma, and the SVG title and desc elements that screen readers depend on. On the text side, safe mode preserves the zero-width joiners that build compound emoji and the non-joiners that Arabic and Indic scripts require.

Files are sent to a serverless function, parsed in memory and discarded when the response ends. Nothing is written to disk or object storage, there is no database and no logging of file contents. Text is never sent anywhere at all — the entire text analysis runs inside your browser. The one trade-off is a 4 MB limit on file uploads, which comes from the serverless request body cap rather than from the parsers themselves.