WebMCP

Why we included this project

WebMCP tackles a genuinely awkward problem: getting an AI assistant to do something inside a web app instead of just reading it. It defines a protocol that exposes functions running in a page's JavaScript to LLMs as standard MCP tools, so an agent can trigger real actions, like managing tasks in the included todo demo, while operating within the user's session. The repo works well as a reference because it ships demo code in vanilla TypeScript and React, plus a script-tag drop-in for adding it to any page, with a Vue version contributed by the community. One honest caveat: this repository is a historical snapshot. The API was accepted as a W3C deliverable and the open transports moved to a separate WebMCP organization, so use this as a clear explanation of how browser-to-agent tool wiring works, and follow the current standardization effort if you plan to build on it.