For thirty years the web has been built for one kind of visitor: a human with eyes, a cursor, and the patience to read a page. That assumption is now under pressure from AI agents that browse, click, and fill forms on a person's behalf, and they are bad at it, because they are forced to guess their way through interfaces designed for people. At Google I/O this year, Chrome proposed a fix called WebMCP, an open web standard that lets a website expose structured tools, essentially labeled functions and forms, directly to browser-based agents. It is a small technical idea with a large implication: the web may be about to get a second interface, one built for machines.
What WebMCP actually is
WebMCP lets a site describe what an agent can do on it in a structured, machine-readable way, instead of making the agent reverse-engineer the page's buttons and fields. A retailer could expose a "search products" tool and an "add to cart" tool; a booking site could expose "check availability" and "reserve." The agent reads those declarations and calls them directly, the way one program calls another, rather than simulating mouse clicks and hoping the layout has not changed. Crucially, Google is not developing this alone. It is working on WebMCP with Microsoft through the W3C, the body that standardizes the web, with the explicit goal of an open standard that any browser can adopt rather than a Chrome-only feature.
RelatedFirefox Halves Its Release Cycle to Two Weeks
Why the current approach is so fragile
To understand why this matters, look at how agents browse today. They take a screenshot or read the raw page structure, infer where the search box is, type into it, guess which button submits, and pray the site did not move things since the last time. It is brittle, slow, and breaks the instant a site ships a redesign. It is also a security headache, because an agent interpreting a whole page can be misled by hidden text or deceptive elements designed to trick it. WebMCP replaces guesswork with a contract. The site says, in effect, "here are the exact actions I support and how to call them," and the agent uses that interface instead of scraping. The difference is the gap between a program reading documented commands and a program staring at pixels trying to figure out what a button does.
The deeper shift most coverage misses
The part that gets lost in the technical detail is what WebMCP implies about the web's future. For decades the page was the product: businesses spent enormous effort on layout, design, and persuasion because the page was where humans made decisions. If agents become a primary way people interact with services, the page stops being the only thing that matters, and the structured tool interface behind it becomes just as important. A business will need to think not only about how its site looks to a person but about what capabilities it exposes to a machine acting for that person. That is a genuine change in what it means to have a presence on the web, and it raises hard questions about who controls the interaction, who sees the ads, and who owns the relationship with the customer when an agent sits in the middle.
The split among browser makers
Not everyone is rushing in, and the disagreement is the most interesting part of the story. While Google and Microsoft push agentic standards and ship agent features, Vivaldi, Firefox, and Safari have each held back from shipping autonomous agentic features, citing privacy and user control. That divide reflects a real philosophical fork. One camp sees agents acting on the web as the obvious next step and wants to standardize it so it happens in the open. The other worries about what it means to let software take actions on your behalf across the sites you use, what it does to consent, and how much trust you are placing in the agent and the company behind it. Both positions are defensible, and the fact that a standard is being proposed does not guarantee universal adoption.
RelatedThe Anti-AI Browser Backlash Is Fueling DuckDuckGo's Surge
Who this affects
Website owners face a new question: do you expose tools to agents, and on what terms? Doing so could make your service far more useful to the growing population of agent users; refusing could make you invisible to them. Developers get a cleaner, more reliable way to build agent integrations than screen-scraping. And ordinary users stand to benefit from agents that actually work, that can book, buy, and search without fumbling, if the privacy and control concerns the holdout browsers are raising get answered rather than waved away. The technology is arriving regardless. Whether it arrives in a way that respects users depends on those concerns being taken seriously.
Our take
WebMCP is one of those quietly consequential proposals that sounds like plumbing and is actually a redefinition. Building it as an open W3C standard with two rival giants collaborating is the right instinct, because the worst outcome would be every AI company inventing its own incompatible way to drive websites and fragmenting the web into agent silos. But the holdouts are not being obstinate. The questions Firefox and Safari are raising about privacy, consent, and control are the questions that determine whether the agentic web serves users or quietly hands more of their decisions to whoever builds the most aggressive agent. The standard is the easy part. Getting the trust model right is the work that actually matters, and it is barely started.
Source: Chrome for Developers, analysis by GenZTech.