WebMCP
Server-side MCP lets agents call your backend. WebMCP extends the same idea into the page itself, so an agent can submit a form or start a checkout without screenshotting your screen and guessing. It is early, and it is where this is going.
Today, an agent operating in a browser mostly has to look at a rendered page and infer how to use it, clicking buttons and filling fields the way a person would, with all the fragility that implies. WebMCP replaces the guessing. A site exposes named, described tools, and the agent calls them directly, the same pattern as the Model Context Protocol, brought into the page.
It can be declarative, through attributes on your HTML, or imperative, through a JavaScript interface for richer apps. Either way, the agent stops puppeteering your UI and starts using a real interface.
Why it matters even though it is early
WebMCP is still maturing in browsers and is not yet a production default. So why care now? Because the brands that learn it early will be the ones agents can operate first, and that head start matters most in agentic commerce, where the agent that can complete a checkout cleanly wins the sale. Get the readable and citable layers right first, then move into being callable. The whole progression is laid out in agent legibility.