benchr
Here is the problem WebMCP is reacting to. When you tell an agent "find a hotel in Istanbul and book it," the agent lands on a website built for human eyes and has to reverse-engineer it: which element is the search box, which button is "Book Now," what the date picker expects. That works, sort of, and breaks constantly. It is brittle because the agent is guessing at intent from pixels and DOM nodes.
What WebMCP proposes
Instead of guessing, the agent reads a contract. The mechanics, per the Chrome and W3C materials:
- A browser API,
navigator.modelContext, through which a site declares its capabilities to agents running in the browser. - A "Tool Contract" — a structured list where the site states, in effect, "here is what you can do here," each entry a callable function with defined inputs.
- Two styles of API. A declarative path for standard actions that can be described directly in HTML forms, and an imperative path, using JavaScript, for more dynamic interactions.
The standard is being developed in the W3C's Web Machine Learning Community Group, created by engineers at Google and Microsoft, and was published as a W3C Draft Community Group Report earlier in 2026. It is available as an early preview in Chrome, with Gemini in Chrome slated to support the APIs. Analysts have described the ambition as the "USB-C of agent–web interactions" — one standard interface replacing a tangle of bespoke scraping — which is a framing from commentary, not an official claim.
Why call it a warning
Because it points at the same shift the rest of this season's news does: the web is gaining a second audience. For two decades you built pages for humans, with SEO as the concession to machines. The agentic web adds a reader that does not look at your layout at all — it wants to call functions. WebMCP is the most explicit statement yet that "agent-readable" is becoming a first-class property of a website, alongside "human-readable."
If that direction holds, the sites that win agent traffic will be the ones that expose clean, reliable tools — and the ones that rely on visually clever, hard-to-parse interfaces will be the ones agents skip or fumble. It is structurally similar to the earlier transition where structured data and clean markup started mattering for search; this is the same idea aimed at action rather than retrieval. Our tools section and the model index track the broader agent ecosystem this plugs into.
What it means for developers and site owners
- Do not ship to it yet — but do inventory it. WebMCP is early preview. The useful move now is to list the three to five actions a user would most want an agent to perform on your site (search, add to cart, book, check status). Those are your future tools.
- Notice you may already be close. The declarative path leans on HTML forms. Clean, semantic forms with clear names are not just good accessibility practice — they are the raw material a contract like this would expose.
- Treat it as the action-layer sibling of getting cited. Being quotable wins you AI-answer visibility; being callable wins you agent actions. Both are about being legible to machines — see how to get cited inside AI answers for the retrieval half of the same discipline.
- Plan for the governance questions. Exposing functions to agents raises real issues — authentication, rate limits, abuse, and what an agent is allowed to do without a human confirming. A serious adoption plan budgets for those, not just for declaring tools.
Honest caveat: proposed web standards do not all ship, and the ones that do can change shape on the way. Browser-version specifics in early coverage have varied, and multi-vendor support is not guaranteed by a community-group report. The signal here is the direction — machine-callable websites — more than any particular API surface as written today.