WebMCP in Chrome 149: Google's Origin Trial Lets Websites Serve AI Agents

Google launched a public origin trial of WebMCP starting with Chrome 149, letting developers expose website actions to AI agents as structured, callable tools on live sites — the first time the draft standard has been testable outside a lab flag. Microsoft's Edge joined the trial with Edge 150, and spec editors from Google and Microsoft are developing the proposal inside the W3C Web Machine Learning Community Group.

The problem WebMCP (Web Model Context Protocol) addresses is concrete: today, most AI agents interact with websites by taking screenshots, parsing raw HTML, and guessing what each button does. A pop-up can derail the whole sequence. It is slow, expensive, and unreliable. WebMCP lets a webpage declare its features as tools an agent can call directly — effectively turning the site into an API without the site maintaining a separate one.

Consider a flight search from Newark to Austin on Nov. 6. Without WebMCP, an agent must work out the interface, field by field, screenshot by screenshot. With WebMCP, the site exposes a "searchFlights" tool that specifies origin, destination, date, and passenger count; the agent supplies the parameters and receives matching flights, replacing multiple browser interactions with one structured call that cuts time, tokens, and processing.

What does WebMCP change for developers?

Two APIs are specified. The Declarative API converts existing HTML forms into tools by adding attributes: toolname and tooldescription (both required), plus optional toolautosubmit and toolparamdescription. A restaurant reservation form tagged with toolname="reserve_table" becomes agent-callable, and the browser auto-generates a JSON Schema from the form fields describing what the agent must provide.

The Imperative API, for dynamic interactions, registers tools in JavaScript through the browser's document.modelContext interface with four properties: name, description, input schema, and an execute function. Tools live only in an open tab; navigation or closing the tab revokes access.

How are consent and security handled?

The spec builds in human oversight. Forms default to manual submission unless a site opts into toolautosubmit. A tool's code can call requestUserInteraction to pause and confirm sensitive actions such as purchases. WebMCP must run on HTTPS pages under permissions-policy controls, and a consequentialHint annotation flags risky actions.

The trade-off is real: tools can act inside a user's logged-in session, and prompt injection (malicious instructions hidden in content such as a product review) could trick an agent into unauthorized actions. The spec's safeguards do not guarantee safe behavior, so developers should limit tool scope, validate permissions, and require confirmation before purchases, account changes, or deletions.

Which browsers support it?

Browser positions, per vendor documentation, split sharply:

• Chrome: public origin trial from Chrome 149, off by default; local testing behind chrome://flags/#enable-webmcp-testing
• Edge: origin trial from Edge 150; Microsoft engineers co-edit the spec
• Firefox: no implementation; Mozilla's official position is neutral
• Safari: no implementation; Apple's WebKit team officially opposes the specification

These are confirmed vendor positions, not tool-signal speculation. Adoption signals elsewhere are early but real: ChatGPT's desktop app supports WebMCP through site tools in its built-in browser for some accounts; Google has announced plans to support it in Gemini in Chrome; and Expedia, Booking.com, and Target are experimenting with it.

Should marketers and SEOs act now?

The practical question is what this changes for site owners. The answer today: groundwork, not implementation. WebMCP is still evolving — the API surface shifted from navigator.modelContext to document.modelContext, so anything built now may need rework. A sensible path is a small-scale test via the origin trial (register a domain, add a trial token) or the local flag, paired with the Model Context Tool Inspector extension from the Chrome Web Store.

The stakes are visible in the agentic-browser market: Chrome Auto Browse (Jan 2026) uses Gemini to complete multi-step tasks; Perplexity's Comet (Jul 2025) and Opera Neon (Sep 2025) both deploy agents that navigate sites and fill forms. As agents take on more user tasks, how easily an agent can use a site could become a revenue factor — and WebMCP adoption may mirror responsive design's trajectory after mobile browsing took off.

Impact spans verticals. In ecommerce, agents could chain "searchProducts," "checkPrice," and "addToCart" tools. Travel sites could expose "searchAvailability" and "bookReservation." B2B SaaS dashboards could offer "createReport" and "exportData" from logged-in sessions; support sites could expose "createTicket."

WebMCP is complementary to existing protocols, not a replacement. Traditional MCP (Model Context Protocol) connects AI systems to tools via servers; WebMCP runs browser-native in an active tab and inherits the user's session state. OpenAPI describes server-side API endpoints; llms.txt tells AI systems what to read, while WebMCP defines what an agent can do.

The immediate, low-cost step: audit your five to ten most valuable user actions — lead forms, product search, booking, checkout, support — and ensure clear labels, properly defined inputs, and reliable validation. Well-structured forms lower WebMCP adoption costs later, and Lighthouse's experimental Agentic Browsing audit already checks whether registered tools work correctly. Watch whether Mozilla moves off its neutral stance and how widely Gemini's Chrome support ships — those two signals will tell you when agent readiness stops being optional.