WebMCP: How every website talks to agents — Tara Agyemang, Google Chrome(AI column)
Interesting. speech Tara Agyemang from Google Chrome about WebMCP. Not the emergence of a new browser API, but a more practical idea: normal browser flow management for agents can be arranged not through “look at the screen and guess where to click”, but through explicit tools that the site itself gives to the agent.
I really feel this could be an interesting opportunity. Now the browser agent often does expensive work: reads the DOM, watches the accessibility tree, takes a screenshot, guesses the visual structure, calculates coordinates and clicks. And then it breaks, because the banner loaded on top and everything moved out. It’s impressive as multimodal reasoning, but engineering looks like a workaround.
WebMCP offers a different approach. The site announces tools, for example: search_concerts, purchase_ticket, filter_results or submit_application. Each tool has a name, description and schema parameters. The agent does not guess the assignment of the button by pixels, but sees the menu of permissible actions. In the demo, Tara shows this on the maze game and the ticket site: the agent does not “poke” the interface, but consistently calls the page tools, while the UI remains synchronized.
It seems to me that there is an analogy with accessibility features to increase accessibility for people - when the interface is made only for a sighted person with a mouse, another artist has to restore meaning from external features. WebMCP looks like an access layer for agents: the site explains not only what it looks like, but also what actions actually exist in it.
Tara does not sell WebMCP as a replacement for conventional MCP. MCP is usually about server-side integration: the agent goes to a separate backend tool/API. WebMCP is a client-side browser context. You need an open tab or webview, tools live on the page, and the user sees the state and can return control to himself. It’s a human-in-the-loop workflow inside a familiar web UI.
Technically, there are two ways. 1️⃣ Declarative API Suitable for HTML forms: add attributes, browser collects schema from fields 2Imperative API needed for complex flow: Through JavaScript, register the tool and inside execute, invoke the existing application logic. In the Chrome documentation, this is progressive enhancement, not rewriting the site as agents.
The status of the standard is still early, and it is important not to confuse it. WebMCP is described as a proposed web standard. current specification It is published by the Web Machine Learning Community Group as Draft Community Group Report, meaning it is not yet a W3C Standard or W3C Standards Track document. Google announces early preview 10 February 2026 Origin Trial is available with Chrome 149.
If you asked me honestly, I would share two answers. For Chrome implementation, Chrome Status now shows the origin trial on the M149-M156 and the ship stage with the M157 for Desktop, Android and WebView. Chrome switches to two-week release cycle with M153 8 September 2026 The M157 is expected to arrive in early November. 2026 years unless the plan moves. But that doesn’t mean that the “web standard came out in GA” in a broad sense: the specification is still being discussed, the API changes, and the interbrowser history will depend on feedback and implementation by other browsers.
The practical conclusion is as follows: if a product has complex flow - purchase, application, support, booking, diagnostics, multi-step settings - it is worth thinking now which actions in them are "tools", and which are just visual design. Good semantic HTML, accessibility, fast pages and clear states remain the base. WebMCP adds the following layer: a secure and observable control knob instead of the torment of a multimodal model. I think that’s what will distinguish a mature agent-ready platform from a beautiful demo. Not “the agent learned to click like a human,” but “the system gave the agent the right interfaces and left the person in control.”
P.S. Then I will try to deal with these possibilities in my public projects like system-design.space or polomodov.tech:))
#AI #Agents #Web #Software #Architecture #Engineering