Definition
WebMCP is a proposed web standard that lets a web page register structured tools, each with a name, a natural-language description and a JSON Schema for its inputs, so that AI agents in the browser can call them instead of guessing which buttons to press. The specification is a Draft Community Group Report of the W3C Web Machine Learning Community Group, edited by engineers from Microsoft and Google. It is not a W3C Standard and is not on the standards track.
The idea is to let a page act like a Model Context Protocol server whose tools run in client-side JavaScript. Without it, an agent doing computer use reads the DOM or screenshots and simulates clicks and typing, which Chrome's documentation calls actuation. With WebMCP the site declares that a control means checkout or filter_results, and the agent calls that tool with typed arguments while the user watches the same page.
How it works
There are two APIs. The imperative API registers tools in JavaScript through document.modelContext.registerTool() (early previews used navigator.modelContext). The declarative API adds annotations to ordinary HTML forms so a form becomes a tool. Both are gated by a tools Permissions Policy that allows same-origin use by default and blocks cross-origin iframes unless they are given allow="tools". A tool can require the user to confirm a sensitive action such as a purchase.
Chrome lists three limits. WebMCP is designed for a human in the loop rather than headless browsing, complex apps may need refactoring to expose their state, and an agent only learns which tools a site offers after visiting it.
Where it runs (October 2026)
Google opened an early preview behind a flag in Chrome 146 in February 2026, followed by an origin trial starting in Chrome 149. Microsoft runs its own origin trial that started in Edge 150 and is open until March 30, 2027. OpenAI turned on WebMCP in the built-in browser of the ChatGPT desktop app on August 25, 2026 under the name site tools: ChatGPT Work and Codex discover and call a page's tools, though not through ChatGPT's Chrome extension. Brave has experimental support in its Leo assistant.
Browser engines are split. WebKit resolved its standards position as oppose in June 2026, arguing that gaps in agent usability should be fixed in HTML and ARIA semantics, where screen readers and keyboard users benefit too, and that a tool's description cannot guarantee what it does. The spec editors say they are working with Mozilla on API changes to win its support.
Why it matters for AI search visibility
WebMCP does not affect whether a page is crawled, indexed or cited. It matters one step later, when an AI agent has already arrived in an agentic browser and needs to finish a task such as a booking, a catalog search or a support request. A site with clear, well-described tools gives agents a reliable path to completion, which is the goal of agent experience optimization. It complements NLWeb, which gives a site a conversational query endpoint, and server-side MCP, which works without a browser.
A sensible starting point is the declarative API on existing forms. Write tool names and descriptions as carefully as page copy, because the agent picks tools from them, require confirmation for anything irreversible, and test with Chrome's Model Context Tool Inspector extension. Since the spec is experimental and Safari support looks unlikely soon, treat WebMCP as progressive enhancement on top of an accessible HTML interface. The Chrome WebMCP documentation and the draft specification are the primary references, and our guide to preparing your site for AI agents covers the wider checklist.
Examples of WebMCP
- A travel site registers a search_flights tool with origin, destination and date parameters, so an agent in Chrome's origin trial can plan a multi-city trip without misreading the custom date picker.
- A SaaS company annotates its existing support form with the declarative API, and ChatGPT Work in the desktop app's built-in browser maps details from the user's conversation into the right fields.
- A retailer marks its place_order tool as requiring confirmation, so an agent can fill the cart and apply a coupon, but the shopper still approves the purchase in a dialog.
- A developer tests tools with the Model Context Tool Inspector extension, sees the agent choose the wrong tool because two descriptions overlap, and rewrites them with distinct names and examples.
