A relay MCP server for pages that refuse to talk to AI agents. Real Chrome, guarded egress, untrusted-by-default output.
An MCP server that browses on an agent's behalf.
You ask Claude to read a page. The site checks the user agent, sees ClaudeBot, and serves a 403. Or it checks the source address, recognises a model provider's range, and does the same. Or the page is a single-page app and there is nothing in the HTML to read anyway.
Ghost Protocol loads that page in a real Google Chrome on a server that belongs to neither of you, presenting an ordinary Windows 11 Chrome identity, and hands back two things: the page as readable text, and a screenshot of what it looked like. Both are labelled as untrusted the whole way.
The full documentation is in the README.