What is WebMCP, really?

The short version

WebMCP adds a way for a website to tell an AI agent — one running in the visitor's own browser, like Chrome's built-in agent — exactly what it's allowed to do on that page. Search a catalog, fill in a form, check availability: instead of an agent guessing by scraping the page and clicking around (slow, fragile, easy to get wrong), the site declares a small set of tools with real names, descriptions, and structured inputs, and the agent calls them directly.

Two ways a site can do this

WebMCP actually has two registration surfaces, and it's worth knowing both exist:

One detail worth flagging since even technical coverage of WebMCP gets this wrong: the JavaScript object moved from an earlier navigator.modelContext to document.modelContext in a May 2026 spec revision. Chrome keeps the old name working as a deprecated alias, but document.modelContext is the current, correct surface. This is exactly why we link primary sources on this page instead of asking you to trust a summary — a fast-moving spec is easy to describe out of date.

Who's actually building it

WebMCP is developed under the W3C Web Machine Learning Community Group — the same open standards process that produced other web platform features. The specification's editors are engineers from both Microsoft and Google:

• Brandon Walderman, Leo Lee, Andrew Nolan — Microsoft
• David Bokan, Khushal Sagar, Hannah Van Opstal, Dominic Farolino — Google

Mozilla and Apple are also participants in the working process, even though Chrome is currently the only browser actually shipping an implementation. That combination — two competing browser vendors co-authoring the spec, developed in the open, with the other major engines at the table — is a genuinely different starting position than a single company's proprietary API. It's also exactly why this isn't being treated as settled: the spec is still a draft, and details are still changing (see the surface-name change above).

Where it stands right now

As of this writing, WebMCP is running as a Chrome origin trial (Chrome 149 through 156, ending mid-November 2026) — meaning real sites can enable it for real visitors with a registered trial token, not just developers flipping a local flag. That's a meaningfully more mature stage than "experimental behind a flag," but it's still pre-standard: no other browser ships it yet, and the API can still change before it's finalized.

Why this matters for your site's future

Browser-native AI agents are already visiting websites today — they just do it the hard way, by reading the rendered page and guessing at what a button or field means, the same fragile approach any browser-automation script has always used. That breaks constantly, and it gives a site no way to say "here's what's safe to do, and here's what needs a human to confirm first."

WebMCP is the alternative: the site itself defines what an agent can do, in a form the agent doesn't have to guess at. Sites that adopt it early get agents that work reliably on their pages instead of misfiring; sites that don't will simply keep being scraped, with no say in how. That's the entire reason this index exists — to make "how ready is my site for this" a real, checkable, un-gameable number instead of a guess.

Primary sources — go verify this yourself

• github.com/webmachinelearning/webmcp — the spec source, editors, and changelog
• webmachinelearning.github.io/webmcp — the rendered specification
• w3.org/community/webmachinelearning — the official W3C Community Group page
• developer.chrome.com/docs/ai/webmcp — Chrome's own implementation docs (imperative + declarative API references)
• developer.chrome.com/blog/ai-webmcp-origin-trial — how to join the origin trial