🎧 Prefer to listen?

A few weeks ago at Google I/O 2026, between the Gemini 3.5 Flash demo and the smart glasses reveal, someone said the words “WebMCP” on stage. Most people missed it. I didn’t.

Because this one changes everything about how AI interacts with the web — and nobody’s explaining it in plain English.

The problem WebMCP solves

Right now, when you ask an AI assistant to do something on a website — book a flight, add something to your cart, fill out a form — it has to basically guess what to do. It looks at the buttons and links on the page and tries to figure out which one is “Book Now” and which one is “Add to Favorites.”

This is called screen scraping, and it’s terrible. It’s slow. It breaks whenever a website changes its layout. And it makes mistakes — the kind of mistakes where you end up booking three hotel rooms instead of one.

I wrote about how AI currently calls tools a few weeks ago. The short version: AI models can connect to external services through APIs and structured tool calls. But websites? Websites were the missing piece. Until now.

What WebMCP actually is

WebMCP stands for Web Model Context Protocol. It’s a proposed open web standard — built by Google and Microsoft engineers under the W3C — that lets websites tell AI agents exactly what they can do.

Think of it like this: every website gets a menu. Not a visual menu you click through, but a structured list that AI agents can read. The menu says “here are the things you can do on this site” — search for flights, add to cart, book a demo, fill out a contact form. Each item on the menu comes with clear instructions: what information you need to send, and what you’ll get back.

The AI agent reads the menu, picks the right item, sends the correct information, and gets a structured result. No guessing. No clicking around. No breaking when someone redesigns the homepage.

How is this different from what we have now?

If you’ve used Gemini in Chrome or any browser automation tool, you know the current experience is hit-or-miss. Sometimes it works. Sometimes it clicks the wrong thing. Sometimes it fills out a form incorrectly and you have to redo everything manually.

WebMCP replaces that guesswork with certainty. Instead of the AI saying “I think this button does what you want,” the website says “this function does exactly what you want, and here’s how to call it.”

The difference is like the difference between giving someone directions by saying “look for the big tree and turn left” versus giving them GPS coordinates. One works when conditions are perfect. The other works every time.

What does this mean for regular people?

Here’s where it gets interesting for those of us who aren’t web developers.

Shopping gets smarter. Imagine telling your AI assistant “find me running shoes under $80 in my size and add them to my cart.” With WebMCP, the shopping site’s AI menu includes an “add to cart” function that takes your size, budget, and preferences as inputs. The agent calls it directly. No clicking through filters. No entering your size three different times.

Booking travel gets simpler. Instead of navigating through seven screens to book a flight, your AI agent calls the airline’s “book flight” function with your dates, preferences, and payment info. The whole thing happens in one step.

Filling out forms becomes automatic. Every contact form, application, or signup page that supports WebMCP becomes something your AI can fill out correctly on the first try. No more “please enter your phone number in the format (XXX) XXX-XXXX” errors.

AI search results get better. When AI-powered search engines recommend websites, they’ll prioritize sites with WebMCP support — because those sites let the AI actually complete tasks, not just provide links. If you run a website, this matters for your visibility in AI search results.

The three-protocol stack

Google didn’t just announce WebMCP in isolation. It’s part of a bigger picture:

  • MCP (Model Context Protocol) — handles connections between AI and your tools, databases, and APIs. I covered this in how AI calls other tools.
  • A2A (Agent-to-Agent) — lets different AI agents talk to each other and coordinate tasks.
  • WebMCP — handles the browser layer. AI agents interacting with websites.

Together, these three protocols answer the question “how does an AI agent actually do things in the real world?” MCP handles the backend. A2A handles agent coordination. WebMCP handles the web.

Should you care right now?

Honestly? Not yet — but soon.

WebMCP is in Chrome 149 origin trial right now. That means developers can start testing it, but it’s not available to regular users yet. The standard is still being finalized through the W3C.

But here’s what I’d watch for:

Browser support. Chrome is first. If Microsoft and other browsers adopt it (and the W3C backing suggests they will), this becomes a web standard, not a Chrome feature.

Website adoption. The websites that adopt WebMCP early will get priority in AI-powered browsing and search. If you run an online business, this is something to start thinking about.

AI assistant support. Once ChatGPT, Gemini, and other AI assistants support WebMCP, the experience of using AI to browse and shop will go from “sometimes works” to “just works.”

What I’m watching next

A few things I’m paying attention to as this moves forward:

Security. If websites can define what AI agents can do, what stops a malicious site from registering fake tools? The security model needs to be airtight before this touches anything involving payments or personal data.

Cross-browser adoption. Chrome-first is fine for now, but this only becomes transformative when it works everywhere. The W3C backing is promising, but standards processes are slow.

The long tail. Big e-commerce sites will adopt this first. But the real magic happens when every small business website, every booking page, every contact form supports WebMCP. That’s years away, but it’s the direction.

If you want to understand more about how AI agents work under the hood, check out my breakdown of AI orchestrators and how AI calls other tools. WebMCP is the next piece of that puzzle.

I test AI tools so you don’t have to. Want to see what else is changing? Check out the tools I actually use every day or start here if you’re new to all this.