Google's WebMCP Wants AI Agents to Stop Guessing at Your Website

Google and Microsoft rarely agree on much. This is one of the rare times they do, which alone should tell you something: they're building the same fix for something every AI agent gets wrong. It doesn't actually know how to use your website.

Ask an agent to buy something for you today and watch what it actually does. It takes a screenshot of the page. It looks for something that resembles a button. It clicks, waits, and screenshots again, reading your screen the way you would, except slower, and you're paying for every look.

That works well enough to demo. It breaks the moment a site changes its layout.

Here's what makes that stranger than it first sounds: the website already knows exactly what it can do. It has a search, a cart, a checkout, a booking flow. None of that is written down anywhere a program can read. All of it is buried inside a layout built for human eyes, not code. The problem was never that agents are bad at reading pages. It's that pages were never written for anything except people.

What It Is

In plain terms, WebMCP is a way for a website to hand an AI agent a proper menu of what it can do. Search, add to cart, book a slot: each action gets a name, a plain-English description, and a list of exactly what information it needs. Instead of an agent guessing from a picture of your page, the site simply tells it, the same way a restaurant menu tells you what's on offer without you having to guess from the smell of the kitchen.

It's built by the Chrome and Edge teams (Google and Microsoft, through the W3C's Web Machine Learning Community Group, not one company's pet project), and it's live in Chrome 149 as an origin trial as of this year.

And the companies already testing it aren't small: Expedia, Booking.com, Shopify, Etsy, Target, Instacart, Redfin, Credit Karma, TurboTax. That's most of e-commerce and half of consumer finance signalling, at once, that they think this matters.

It helps to see where this sits among the other ways an AI agent can reach a website today, because there are more than people assume. An agent can call a company's backend directly if it has an API key: precise, but the site itself is never involved, and someone has to manage that key. It can connect to a server the company built specially for agents, which is better because someone who understands the product defined the actions, but the page a customer actually sees is still skipped entirely. It can do what most agents do today and just look at the screen, clicking around like a very literal-minded human. Nothing to set up, but slow, and one redesign away from getting lost. It can read a page's underlying code instead of a picture of it, which is more reliable than pixels but still leaves it guessing what an unlabelled button is for. Or the company can ship its own built-in assistant. It works, but it isn't your agent: the vendor picks the model, pays for it, and nothing it learns carries anywhere else.

WebMCP is the option that gives up the least. Your own agent still does the work, there's nothing extra to configure, and for the first time it's being told what things do instead of being left to guess.

There's a genuinely clever part most explanations skip. A WebMCP action runs inside the visitor's own browser tab, in their own session, so there's no separate login for the agent to manage and no API key floating around. And the menu of actions changes automatically with who's looking: a visitor who isn't logged in gets read-only actions like search and browsing; the moment they sign in, checkout and order history appear too. The business doesn't build one experience for people and a second one for agents. It's the same site, telling the truth to whoever (human or AI) is currently allowed to see it.

What You Should Know

None of that is deployment.

Being in an origin trial means a company is testing the waters, not that its site has working WebMCP tools live today. As of the most recent independent checks, actual working implementations outside demo pages are close to zero. There are now more tools built to check whether a site supports WebMCP than there are sites that actually do, which tells you plenty on its own.

The bigger tell: no mainstream AI agent calls these tools yet either. Not Claude. Not ChatGPT. Not Gemini. Every one of them is still doing exactly what I described at the start, clicking around blind, on the same web that's supposedly ready for them. Right now, only the browser's own built-in agent can actually call a site's WebMCP tools. Nobody has wired the popular AI assistants up to use them yet.

So the honest state of WebMCP in 2026 is this: the standard exists, the browser support exists, some of the world's biggest websites are testing it, and almost nobody, on either side, sites or agents, is actually using it for anything real yet.

What It Means

It's tempting to read "nobody's using it yet" as "ignore it for now." That's the wrong read.

A gap between a standard existing and a standard being used is exactly the window where being early is cheap and being late is expensive. Once agentic browsing genuinely works, once asking an AI to "book the cheapest flight to Nairobi next Friday" reliably works instead of watching it fumble through a checkout form, the businesses whose sites already speak that language get found first. The ones that don't become the equivalent of a shop with no signage in a city that's stopped reading signs and started asking for directions instead.

Chrome's rollout timeline points to broader browser support through 2026 and formal standardisation sometime in 2027. That's not far off. It's close enough that a business building or rebuilding a website this year should already be asking whether it's structured in a way that makes adding this cheap later, rather than expensive.

So what should you actually do about it? Not a full WebMCP implementation. The spec is still moving, and building against a standard that could change under you is a waste of a client's money.

What's worth doing now is the groundwork that pays off regardless of how the spec settles: a website whose actions (search, book, checkout, filter) are already clean, well-defined operations behind the scenes rather than tangled page logic, and content and product data structured well enough that an agent, or a search engine for that matter, can actually understand what it's looking at. For a site that's already built this way, turning an existing form into something an agent can use is a small addition, not a rebuild. Closer to labelling a door than knocking down a wall.

That's infrastructure work, not speculation. It pays for itself in the ordinary web today and puts you first in line for the agentic one.

For most of the web's history, the only visitor worth designing for was a person. That's starting to change, and the site that tells an agent what it can do will get cleaner, more reliable results than the site that leaves it guessing from pixels.

The web is being rebuilt for agents.

Most sites aren't ready to be found by them yet.

By Nasif K Balinda