There is a local MCP server, a command-line tool and a Claude Skill. All three run the exact cleaner.js this site loads in your browser — the same file, byte for byte — so an agent can inspect and clean text and images without anything being pasted into a web page. The server speaks MCP over standard input and output only, so it opens no network connection.
Four tools, in the order you would actually use them:
inspect_text — lists every invisible or bidirectional character in a string, with its code point and position. Read-only.clean_text — returns the string with those characters removed, plus a count of what went.inspect_image — reports which metadata blocks a JPEG or PNG holds: EXIF, XMP, IPTC, C2PA. Read-only.clean_image — writes a cleaned copy. Never overwrites the original unless you say so.The package is scrubwatermark-mcp. It has zero runtime dependencies, so there is nothing to install before it runs — Node 18 or newer is the whole requirement.
Download the package straight from this site. There is no registry account involved, so this works today:
npm install -g https://scrubwatermark.com/ai/scrubwatermark-mcp-1.0.0.tgz
That gives you two commands, scrubwatermark and scrubwatermark-mcp. If you would rather not install anything globally, run the server file directly with node from an unpacked copy.
claude mcp add scrubwatermark -- scrubwatermark-mcp
Add this to claude_desktop_config.json, which lives in %APPDATA%\Claude\ on Windows and ~/Library/Application Support/Claude/ on macOS:
{
"mcpServers": {
"scrubwatermark": {
"command": "scrubwatermark-mcp"
}
}
}
Any other MCP client works too. The server answers protocol versions 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05, and replies with whichever the client asks for.
The text commands read a file, or standard input when you give no file, and write the cleaned result to standard output unless you pass -o:
scrubwatermark inspect-text notes.txt scrubwatermark clean-text notes.txt -o notes.clean.txt scrubwatermark inspect-image photo.jpg scrubwatermark clean-image photo.jpg -o photo.clean.jpg
Because the text commands stream, they pipe into the rest of your tooling. On macOS that is a one-liner from clipboard to clipboard:
pbpaste | scrubwatermark clean-text | pbcopy
The package also ships a Claude Skill. It teaches the agent the distinction that matters — which layer of a watermark is stored in the file and which is not — so the agent explains the limit instead of promising a removal that cannot happen.
cp -r skill/scrubwatermark ~/.claude/skills/
This site also publishes a machine-readable description of itself for agents that are browsing rather than installed: the site facts skill, listed in the agent catalog.
It cannot remove a statistical AI watermark, and no amount of packaging changes that. That watermark is a bias in how a model chose its words; it is not a value stored in your file, so there is nothing on disk to delete. The server's own tool descriptions say this, which means an agent reading them before it calls anything gets told the truth up front rather than after.
It also does not rewrite wording, and it will not make generated text pass an AI detector. If that is what you need, this is the wrong tool.
No. It speaks only over standard input and output, opens no socket, and reads and writes only the paths you name. That is the same promise the browser tool makes, and it is the reason the server is a local process rather than a hosted endpoint — a hosted one would have to receive your file.
No, and it says so in its own tool descriptions. Claude's watermark is a statistical pattern in how the model chose its words, not a value stored in the file, so there is nothing on disk for a tool to delete. Anthropic's documentation states that nothing is added to the text and that there are no hidden characters. What the server does remove is the layer that really is bytes: invisible Unicode characters and image metadata.
No. The package has zero runtime dependencies, so it runs on Node 18 or newer with nothing to install first. The protocol library is a development-only dependency used to test the server against the official MCP client, and the server never loads it.
No. The compressed image data is copied across verbatim, so the file is not re-encoded and no pixel changes. The one EXIF tag that is deliberately rebuilt is Orientation, because a photo shot upright often stores landscape pixels plus Orientation=6 — delete that tag and the picture displays sideways, which is a visible change even though not one pixel moved.