WebMCP just moved from a behind-a-flag prototype into a public Origin Trial in Chrome, the Chrome team announced, running from Chrome 149 through 156. It is a proposed open web standard for the agentic web: a website declares structured tools that an in-browser AI agent can call directly, instead of the agent having to look at the rendered page and guess. There are two ways to declare them, an Imperative API for defining tools in plain JavaScript, and a Declarative API that turns annotated standard HTML forms into tools. The model-facing surface is small and legible: document.modelContext.getTools() lists what a page offers, executeTool() runs one, and each tool carries annotations like readOnlyHint and untrustedContentHint so the agent knows what it is touching. The companion documentation went up in May and the trial was unveiled at Google I/O 2026.
Why it matters is the contrast with how browser agents work today. The current default is the vision loop: screenshot the page, locate a button by pixels, click, screenshot again, hope the layout did not shift. WebMCP replaces that with a typed function call. The Chrome team frames the gain as faster (one call versus a screenshot loop, with early third-party benchmarks citing 8 to 12 times faster end-to-end task completion on WebMCP-enabled pages), more reliable (no pixel-hunting to break when the CSS changes), and safer (the agent cannot invoke a tool the site did not declare). New in this trial specifically: listing and executing tools, a permissions policy, and cross-origin iframe tool support, the plumbing a real site needs before it can expose anything beyond a demo.
This is the agent-runtime contest reaching down into the web page itself. The fight we have been tracking over who hosts and orchestrates agents now has a front at the document level: the site stops being a thing to be scraped and becomes a typed tool surface, an API for agents whether or not it ever shipped one for developers. If WebMCP standardizes, the agentic web stops being a guessing game. The honest catch is that a web standard with one consumer is not yet a standard: today the only agent that reads WebMCP tools is Gemini in Chrome, Google's in-browser assistant, so adoption and neutrality are both still open questions. The concrete builder move is small and cheap: register for the origin trial, declare your two or three highest-value actions as tools, and measure whether the reliability gain actually shows up against the vision baseline.