At I/O 2026, Google introduced WebMCP—a proposed open web standard that lets sites expose JavaScript functions and HTML forms as structured tools that browser-based AI agents can invoke directly. The experimental origin trial begins in Chrome 149, with Gemini in Chrome integration coming "soon." For developers, it's the first concrete proposal for an agent layer that lives in the browser instead of being scraped off the DOM.
What WebMCP Actually Proposes
The problem WebMCP attacks is the one every browser agent has been brute-forcing for two years: an agent loading a page has to read pixels and DOM nodes, guess what the UI is, plan a click sequence, hope nothing shifts, and pray the form submits. It is unreliable, slow, and breaks the moment a designer renames a class.
WebMCP flips that. Sites declare which JavaScript functions and HTML forms they want to expose as tools, complete with structured schemas that describe the inputs, outputs, and side effects. A browser-based agent—Gemini in Chrome or any third-party agent that adopts the spec—queries the page for its tool manifest, picks the right call, and invokes it directly. No pixel parsing, no fragile selectors, no race against React re-renders.
The MCP Spec Lineage
WebMCP is explicitly an extension of the Model Context Protocol, the standard that's quietly become the lingua franca for connecting models to tools. Anthropic introduced MCP for server-side tools; Cloudflare, OpenAI, GitHub, and Google have all shipped MCP servers; WordPress.com added 19 write capabilities to its MCP integration in March. WebMCP brings the same shape into the browser, where the "server" is the web page itself and the "client" is whatever agent has been granted permission to act inside it.
The reuse of the MCP schema is the part that should matter most to web developers. The same tool definition can live in a server-side MCP server and a browser-side WebMCP exposure, so agents acting on behalf of users don't need a separate vocabulary for "things that run on a backend" vs. "things that run in the page I'm looking at."
Why a Web Standard, Not a Library
Google could have shipped this as a Chrome-only library. Doing it as a proposed open web standard via an origin trial is a different bet: it's an invitation for Firefox, Safari, Brave, and every alternative browser maker to implement the same primitive so agents can act consistently across the open web. The pitch parallels how Service Workers, Web Components, and the Fetch API rolled out—Chrome ships first, the spec stabilizes through feedback, other vendors implement.
It also lines up with the broader I/O 2026 framing around "Modern Web Guidance," a new vetted skill bundle of more than 200 use cases for performance, accessibility, and Baseline targeting that ships in Antigravity 2.0 and via npx modern-web-guidance install. WebMCP, Modern Web Guidance, and the new agent DevTools all point at the same thesis: the web platform itself needs first-class affordances for AI agents, not just better libraries on top.
Security and Permissions
The obvious failure mode is a malicious site exposing tools that drain accounts or exfiltrate data. The origin trial framing buys Google time to work through the consent model—what a site can declare, what the user has to approve, what the browser refuses by default, and how a third-party agent embedded in the page proves it has permission to act. These are open questions in the draft spec, and the trial is partly designed to surface answers.
For now, opting in is explicit on both sides. A site has to ship a WebMCP manifest; a user has to grant agent access; a browser has to enforce the boundary. None of this is automatic.
Why It Matters for Web Developers
If WebMCP becomes a real web standard, the long-term implication is significant: your site's UI and your site's agent surface become two facets of the same product. The "Buy" button is a click target for humans and a tool definition for agents. The settings form is a UI for users and a structured endpoint for assistants acting on their behalf. SEO matters; AEO ("agent engine optimization") starts mattering the same way.
In the short term, the Chrome 149 origin trial is a chance to ship a manifest, watch what Gemini in Chrome does with it, and learn what the schema actually wants to be before the spec stabilizes. The companies that ship a usable WebMCP surface in the next few quarters will be the ones agents reach for first when users say "do this on my behalf."