Julian Goldie breaks down the revolutionary Google WebMCP update. This new browser API, co-authored by Google and Microsoft, gives AI agents a 'skeleton key' to websites, moving from fragile scraping to structured tool calls. Discover why this is the biggest shift in SEO since structured data and how it transforms the web from a place we read to a place agents act.
TIMESTAMPS
00:00 Google's WebMCP Update Explained
01:15 Why WebMCP Matters: No More Screenshots
02:40 Google & Microsoft Collaboration
03:50 The Origin Story: Alex Nahas & Amazon
05:10 How It Works: Declarative vs. Imperative APIs
06:30 Human-in-the-Loop & Safety Features
07:45 The Risks: Security & Ad Revenue
09:20 The New SEO: Optimizing for AI Agents
11:00 How to Prepare for the Agentic Web
Full transcript
So we're also going to talk about Google's WebMCP update that literally just came out. So Google just gave AI agents a skeleton key to every website on the internet. On February the 10th, 2026, the Chrome team quietly shipped something called WebMCP into Chrome 146 Canary. It is a new browser API built in collaboration with Microsoft that lets any website hand AI agents a structured menu of everything it can do.
Not a screenshot, not a guess, a direct line to the website's actual functions. Book this flight, search these products, file a support ticket, one API call, done. Now early benchmarks are showing 67% less computational overhead compared to how agents interact with websites today. So task accuracy as well is about 98%.
And the W3C is already moving to make this an official web standard. So that is not an incremental improvement, OK? That is a new layer of the internet being built in real time with Google WebMCP. So let me explain why this matters so much.
Right now when an AI agent wants to do something on a website for you, it's got to pretend to be a human. It takes screenshots of the page, sends those screenshots to a large language model. The model figures out where the buttons are. And then the agent clicks them one at a time.
Think about how absurd that is, right? Thousands of tokens burned just to process a single screenshot. And if the website redesigns even one button, the whole thing breaks. The agent is paralyzed.
It is like asking someone to operate a vending machine at it through a foggy window and poking at it with a stick. That is the current state of the art for AI agents on the web. WebMCP changes this completely. So instead of starting through the foggy window, the website just hands the agent a menu.
It says, here are my functions. Here are the parameters they take. Here is what they return. The agent makes one structured call and gets structured data back.
No screenshots, no guessing, no stick. Kushal Sagar, a staff software engineer on the Chrome team of Google, described it as a USB-C of AI agent interactions. One universal plug that any agent can connect to, replacing the current mess of bespoke scraping strategies and fragile automation scripts. Now, here is the part that makes this a really big deal.
This was not a Google solo project. The specification was co-authored by engineers at both Google and Microsoft. Brandon Waterman, Leo Lee, and Andrew Nolan from the Microsoft Edge team. David Bocan, Kushal Sagar, and Hannah Van Opstel from Google Chrome.
When the two biggest browser vendors on earth co-author a specification and ship working code in parallel, that is not a proposal, my friends. That is a fait accompli. Edge support is all but guaranteed. The W3C is already incubating the spec for its web machine learning community group.
Industry observers expect formal announcements by mid to late 2026. And the origin story of this is wild. A guy named Alex Nahas, a backend engineer at Amazon, was trying to solve a very specific problem. When MCP, Anthropics Model Context Protocol, hit in early 2025, Amazon spun up what amounted to one enormous MCP server with thousands of internal tools crammed into a single context window.
It was, in his words, a mess. But the real headache was authentication. MCP required OAuth 2.1, which basically nobody at Amazon had actually implemented for their internal services. Every service had its own authentication story.
So Alex built a workaround. He realized the browser already handles authentication for all of Amazon's internal services through a federated login system. If you could make the browser itself the bridge between AI agents and web applications, you could just skip the whole backend server requirement entirely. He built a Chrome extension called MCP-B and the underlying protocol he called WebMCP.
Teams at Google and Microsoft had independently been prototyping the same idea. Chrome was working on something they called script tools. Edge was circling the identical question. When they saw what Alex had built, they joined forces through the W3C.
That is three separate engineering teams arriving at the same conclusion independently. That does not happen unless the problem is real and the timing is right. So how does this actually work? Well, there's two APIs for this.
The declarative API handles simple stuff that already exists within HTML forms. If your website is clean, well-structured forms, you're already 80% of the way there. You add some tool names and descriptions to your existing markup and agents can call those forms directly. The imperative API is for the complex stuff, dynamic interactions that require JavaScript.
This is where you register richer tool schemas using a function register tool on a new browser API called navigator.modelContext. So you can expose functions like search products with query and filter parameters or order prints with copies and page size. Full parameter schemas with natural language descriptions running entirely client-side in the browser. In other words, your website's existing front end.
JavaScript becomes the agent interface. So there's no Python server. There's no Node.js backend. And the same code that handles your forms and queries today can now talk directly to AI agents.
And there is a critical redesign decision baked into this that I want to highlight. So WebMCP is explicitly built around human-in-the-loop workflows, not fully autonomous agents running wild on the internet. The specification identifies three pillars. Context, meaning all the data agents need to understand what the user is doing.
Capabilities, meaning the actions an agent can take on your behalf. And coordination, meaning the handoff between you and the agent when it hits something it cannot handle on its own. There is even a built-in function called requestUserInteraction that pauses agent execution and asks for your explicit confirmation before sensitive action. So this is not really about replacing you.
It's about your AI system being able to actually do things on the internet reliably instead of fumbling around taking screenshots. Now, I want to be honest about the skepticism here because there are legitimate concerns. For example, the Decoder pointed out something important. If AI agents handle tasks like product searches, price comparisons, and bookings autonomously, fewer people actually visit websites directly.
That means operators could lose ad revenue, direct customer contact, and the ability to build relationships with their users. The web, as they put it, risks becoming background infrastructure. And there is a security problem that's far from solved. Google's own developer said that protecting against prompt injection attacks is the responsibility of individual AI agents, not the web MCP API itself.
Meanwhile, even leading AI models still fail against targeted prompt attacks more than three out of 10 times. We've all seen that with OpenClaw, for example. Alex Nahas himself flagged what he called the lethal trifecta. Imagine you have two tabs open.
One is your bank, one is a malicious site. Browser agents sitting in the middle treats both as equal sources of context. The malicious tab could theoretically instruct the agent to extract data from the banking tab. That is not a theoretical risk.
It is architecturally inevitable with how browser agents work today. Now, these are real problems that do not have complete solutions just yet. But here's why I think the direction still matters enormously despite those concerns. An SEO expert named Dan Petrovich called webs MCP the biggest shift in technical SEO since structured data.
And that comparison is actually instructive. Think about what happened when search engines needed structured signals to understand websites. We've got sitemaps, robots.txt, canonical tags, schema.org. An entire industry of search engine optimization was born.
Web MCP is the beginning of the exact same paradigm shift except for AI agents instead of search crawlers. If schema.org gave the web its nouns, its structured descriptions of what things are, web MCP gives the web its verbs, its structured descriptions of what things could do. That is a profound distinction. We are watching the web evolve from a place humans read to a place agents act.
And this connects to something much bigger. Companies that add webs, the companies that add web MCP support to their websites first will be discoverable by AI agents. The ones that do not will not even exist in the agent's decision space. So, you know, if you don't adapt now, you're probably gonna get left behind.
And when someone tells their AI assistant to find the cheapest flight to New York on Monday, the airline with web MCP tools will get that booking. The airline without them will not even be considered. That is the exact same dynamic that built the SEO industry two decades. And I get that some people are going to hear this and think, great, another thing that mostly benefits Google and Microsoft.
And that criticism has teeth. Argent from Rover called web MCP a Trojan horse, arguing that Google wants every website to expose APIs so Chrome's agent can serve users directly. You do the work, Google gets a relationship. Fair point.
But the counter argument is also strong. This is a W3 standard, not a proprietary Google protocol. It is magnostic, meaning any agent powered by any AI model can use it. The demo extension runs Gemini 2.5 Flash via API, but the spec is designed so that Claude, GPT, open source models, or whatever comes next can all discover and use these tools.
The browser is a mediator, not any single AI company. So what do you actually do with this information? Well, if you're a web developer, go look at the web MCP explainer on GitHub right now. The repository is at web machine learning, sign up for the Chrome early preview program, start prototyping.
If your app already has clean form structures and well-separated business logic, you might be surprised how close you already are. If you are not technical, but you run a business with any kind of online presence, start inventoring the functions on your website that would be safe and useful to expose as tool contracts. Booking, search, support, ticket creation. Talk to your engineering team about a pilot.
If you are in SEO or digital marketing, like I am, for example, I have an SEO agency, this may be the biggest thing you need to understand right now because tool discoverability is the new indexing problem. Tool descriptions are the new meta descriptions. The quality of your tool names, descriptions, and schemas will directly determine whether an agent selects your site over a competitor's. And if you lead a team, have a real conversation about what the agentic web means for your customer touch points.
Because when AI agents start handling a meaningful percentage of commerce and support interactions, the websites with structured, reliable tools will capture that traffic. Everyone else becomes invisible. The spec is quite early to be fair, you know, and the security model is open questions. The API surface will almost certainly change before it stabilizes.
But Chrome already has working code behind the flag. Microsoft is co-authoring the standard. The W3C is providing institutional scaffolding and engineers from Amazon, Google, and Microsoft all independently arrived at the same architectural conclusion. That is not hype, that is convergence and convergence at this level among competitors, this large tends to signal that, you know, something big is gonna happen.
The question is not whether the agentic web is coming, it is whether you'll be ready for when it arrives. So thanks for watching. If you haven't already, check out the AI Profit Boardroom. This is my AI automation community where we help business owners scale, save time, and grow their business and get more customers using AI automation.
It covers all my best playbooks, a community of 2,300 people, and it's available at the AIProfitBoardroom.com. We also do weekly coaching there four times a week. I update it daily with the newest updates like this and it's available at the AIProfitBoardroom.com. Hope to see you inside there.
Cheers.
More episodes