A page can hand an agent a set of buttons. That is a different question from whether an engine can quote you.

What WebMCP does

A page calls document.modelContext.registerTool() with an object that has three required parts: a name, a description written for a model, and an execute function that does the work. An inputSchema in JSON Schema describes the arguments. The same object carries getTools(), executeTool() and unregisterTool().

A booking form, search box or filter becomes a named function with a described signature, so an agent no longer has to guess at your markup.

How it differs from MCP

MCP, the Model Context Protocol, is a backend protocol: you run a server, it exposes tools, and a model connects to it. WebMCP puts the idea inside the page. The specification describes such a page as an MCP server that implements tools in client-side script.

The tools run in the browser, inside the session the reader is already signed in to, so there are no second credentials and no separate service. It also means the agent acts as the reader, with the reader's access, which makes security a central question.

Reading and acting are separate audits

An engine that cites you is reading. An agent that books your appointment is acting. Our rubric measures reading: whether a site can be retrieved, understood, trusted and quoted.

The two are independent. A site can be quotable with no agent tooling, or expose a dozen working tools and be impossible to cite. So our audit reports what a site declares here and does not score it either way.

Where it stands today

A Draft Community Group Report at the W3C Web Machine Learning Community Group, dated 15 September 2026, which states that it is not a W3C Standard and is not on the W3C Standards Track. Chrome and Edge are running origin trials. ChatGPT Desktop supports it. Firefox and Safari have neither shipped nor refused.

The full implementation table, with the date every line of it was read.

Two common errors

Both confirmed against the specification and Chrome's documentation on 17 September 2026.

• The object is document.modelContext, not navigator.modelContext. Much secondary coverage, including at least one widely read wiki, says navigator, and code written from it will not run.
• The declarative HTML syntax is not specified yet. Several demos annotate ordinary form elements instead of writing JavaScript, but section 4.3 of the specification, where that syntax would be defined, reads in full that it is a TODO.

The adoption figures in circulation have no method

Some claims give a large share of enterprise sites, and a larger share of e-commerce sites, as having implemented WebMCP. We found no methodology attached to those figures anywhere they appear, and the public directory of WebMCP-enabled sites lists none. We do not repeat them.

What we will not tell you

We do not recommend adopting it, and we do not say it will bring agent traffic. Nobody has published usage data that would support that.

Exposing callable tools widens what an agent can do on your site, in the reader's own session, while prompt injection is unsettled. We cannot see your authorisation model from outside, so we cannot tell you it is safe. If you adopt it, do it with your engineers.

We can tell you what your site declares today, beside everything else a machine can read about it. The free scan takes about twenty seconds and asks for nothing.