WebMCP: telling AI agents what your site can do
A proposal from Google and Microsoft engineers lets websites publish actions for AI agents. What it is, how ready it is, and the one thing to do now.
In December we wrote about AI browsers visiting your website: agents in Chrome, Atlas or Comet that read your pages, look at screenshots and click buttons on a customer's behalf. We said they work best on clean, well-labelled sites, because they are essentially guessing what each part of the page does.
There is now a serious effort to remove the guessing. It is called WebMCP, and this month it moved from an idea into something developers can try on real websites.
What WebMCP is
Instead of an agent working out how to use your site by looking at it, your site tells the agent what it can do.
A travel site could declare a tool called "search flights" that takes a departure city, a destination and a date. A clinic could declare "check available appointments" and "book appointment". A shop could declare "search products" and "add to cart". When an agent arrives, it reads that list and calls the actions directly, with structured inputs, instead of hunting for the right box to type in.
There are two ways to do it. The declarative way builds on ordinary HTML forms: you add a little extra description to forms you already have, and they become tools an agent can use. The imperative way uses JavaScript for richer actions such as filtering, configuring a product or anything that changes the page as you go.
The name comes from MCP, the Model Context Protocol, which already lets AI assistants connect to outside software. WebMCP brings the same idea into the web page itself, so the tools live where the user is already signed in and looking.
Who is behind it, and how far along
The proposal was first published last August by engineers from Microsoft and Google, under a W3C community group. Community groups incubate ideas. This is not yet an official web standard.
In February, Chrome added an early preview behind a developer flag. At Google I/O on 19 May, Google announced a public origin trial starting in Chrome 149, which lets real sites switch the feature on for their visitors while it is tested. On 9 June it published the sign-up details. Google also said Gemini in Chrome will soon be able to use WebMCP tools, and showed the logos of brands it says are experimenting, including Expedia, Booking.com, Shopify, Etsy, Instacart, Target and Redfin.
There are caveats. No shipping consumer agent is confirmed to use WebMCP yet; Gemini in Chrome support is "soon". Safari and Firefox have made no commitment. Details of the API may still change. And in May, Google's Lighthouse testing tool added an experimental "agentic browsing" check that includes WebMCP, which shows where Google is heading but is not a ranking signal.
Why it could matter
If it takes off, WebMCP would make agent visits to your site faster and more reliable. An agent calling "book appointment" with a date and a time does not misread a calendar widget or click the wrong slot. You also decide exactly which actions agents can take, rather than leaving them to discover whatever the page allows.
That last point matters for safety too. You can expose search and availability freely, keep booking behind a clear confirmation step, and simply not offer actions you would rather a person did.
What to do now
For most businesses, nothing urgent. This is an experiment, and the agents that would use it are only starting to arrive. But there is a cheap step that pays off either way.
Get your forms right. The declarative half of WebMCP builds on well-made HTML forms, real form elements with clear labels, sensible field names and proper validation. Those same forms already work better for screen readers, browser autofill, today's AI agents and ordinary customers. If your booking or enquiry form is a tangle of custom scripts, fixing it is worth doing this year regardless.
If you run a site where agent traffic could matter soon, such as travel, bookings, retail or anywhere customers compare options, it is reasonable to have a developer try the origin trial on one flow. Pick the action customers ask about most, expose it as a tool, keep any purchase or booking behind a human confirmation, and see how the agents of the next few months use it.
As with llms.txt, be wary of anyone selling WebMCP as a way to rank higher in AI answers. It is about letting agents use your site properly, which is valuable, but it is a different thing.
