Picture an AI agent trying to book a table on a restaurant’s website. Today, it works like a very patient, confused intern: it loads the page, reads raw HTML, guesses which <div> is the date picker, clicks the green button, and breaks when the layout changes. This is screen-scraping and hoping—the automation equivalent of describing screenshots over the phone.

Web MCP proposes something saner. Instead of the agent guessing, your site declares what it can do as clean, structured tools. “Here’s a book_table tool. It takes a date, time, and party size.” No pixel-reading, no guessing. It already runs in Chrome behind a trial, and adding your first tool takes about ten minutes.

The core shift is from scraping to declaring. Today, the agent scrapes the DOM, guesses fields, simulates clicks, and breaks on redesigns. With Web MCP, the page registers a tool, the agent reads its schema, calls it with structured args, and the tool runs real JS—surviving redesigns because the tool is the contract.

Web MCP is a proposed standard by Google and Microsoft in the W3C Web Machine Learning Community Group. It’s a Community Group draft, not finished, and available as a Chrome origin trial from Chrome 149. The proposal lives at github.com/webmachinelearning/webmcp, with Angular support and demo sites. It’s a “try it and shape it” moment.