Most of the AI conversation about websites has been about reading — will ChatGPT quote you, will you show up in an AI Overview. WebMCP is about something different and, for a lot of businesses, more consequential: whether software can use your website on a customer’s behalf.
It’s early. It’s an origin trial. But it’s the first serious attempt to fix a problem that’s been getting worse for two years, and it’s worth understanding before someone tries to sell it to you.
Right now, when an AI agent uses your website, it does it the hard way. It takes a screenshot, works out where things probably are, and simulates clicks and keystrokes.
This is bad for everyone.
It’s brittle — a CSS change, a cookie banner, a layout shift, and the agent fails. It’s slow, because every step is a round trip through a visual model. And it’s adversarial: your site never consented to any of it, has no way to know it’s happening, and no way to offer a better path.
WebMCP inverts that. Instead of the agent reverse-engineering your interface, your website hands the agent a list of the things it can do and the exact parameters each one needs.
Check availability. Get a quote. Search the catalogue. Submit an enquiry. Named operations, with defined inputs, that the agent can call directly.
WebMCP was announced on 10 February 2026, authored by engineers from Google and Microsoft through the W3C Web Machine Learning Community Group. Google shipped an early preview in Chrome Canary, and it’s now in a public origin trial from Chrome 149.
Two things to note about that provenance. First, it’s a genuine standards-track effort at the W3C rather than one vendor’s proprietary hook. Second, Google and Microsoft co-authoring it means the two biggest browser-plus-assistant combinations are pointing the same way — which is about as good a signal as you get this early.
There are two halves to the API.
The declarative half turns ordinary HTML into agent-usable capability. Your existing forms and page elements get marked up so that an agent can see “this is a booking form, these are its fields, this is what each one expects.” No JavaScript required.
The imperative half is a JavaScript API for exposing dynamic functionality as callable tools — the things your site does that aren’t a form. You register a tool, describe it, define its parameters, and the agent can call it.
For most sites the declarative route covers the useful ground. If you have forms that a customer might reasonably want an agent to complete, that’s where you’d start.
One practical wrinkle worth knowing: because imperative tools are registered by JavaScript, timing matters. If your tools register late, Lighthouse’s audit may not catch them, and neither may an agent taking a quick look. Declarative markup doesn’t have that problem.
Since May 2026, Lighthouse’s Agentic Browsing category includes three WebMCP audits:
That last one matters more than it sounds. A malformed schema is worse than no schema — you’ve told an agent you can do something, and then failed to do it. Lighthouse will emit an error rather than a quiet fail.
To get meaningful results from these audits you need Chrome 150 or later and a registered origin trial. If you’re seeing blanks, that’s usually the reason.
Here’s where I’ll be more direct than most people writing about this.
Probably not yet — unless your site does something.
WebMCP is about actions. If your website is essentially a brochure — you tell people who you are, they read it, they ring you — there is very little for an agent to call. Implementing WebMCP would give you a tool list with nothing useful on it, and a passing Lighthouse audit that means nothing commercially.
It’s worth serious consideration if:
It’s not worth it yet if:
That last point deserves emphasis. If your accessibility tree is a mess, WebMCP is the wrong thing to spend money on. Fix the tree first — it’s the substrate agents actually rely on, it helps real people, and it’ll improve your Lighthouse fraction more than a tool registration will.
There’s a parallel track worth knowing about. In January 2026, Google and Shopify launched the Universal Commerce Protocol — an open standard for AI agents to transact with merchants, co-developed with Etsy, Wayfair, Target and Walmart and endorsed by Stripe, Visa, Mastercard and others. Shopify turned on Agentic Storefronts by default for eligible merchants as part of its Winter 2026 edition.
If you’re on Shopify, a lot of this is being handled for you. If you’re running WooCommerce or a bespoke build, it isn’t — and the gap between those two situations is going to widen.
Nothing dramatic. In order:
The standard is a year old at most. The sensible position is to build in a way that makes adopting it cheap later, rather than adopting it now.
Sources: W3C Web Machine Learning Community Group · Chrome for Developers — Lighthouse agentic browsing · Google — agentic commerce tools for retailers
Not sure which category you’re in? Let’s find out where you stand →