What is WebMCP?

The core idea

WebMCP lets a website register "tools" — named, described actions with an input schema — that an AI agent embedded in the browser can call directly, instead of guessing how to fill in a form or click through a UI. It borrows the tool-calling model from the Model Context Protocol (MCP) and brings it to the web page itself.

How it works technically

The proposed API is navigator.modelContext.registerTool(tool, options) (the spec is migrating this to document.modelContext to reflect that tools belong to a specific page). A page calls this at runtime, in JavaScript, to expose one or more tools with a name, a description, and an input schema:

const modelContext = document.modelContext || navigator.modelContext;

modelContext.registerTool({
  name: "check_availability",
  description: "Checks appointment availability for a given date",
  inputSchema: {
    type: "object",
    properties: { date: { type: "string", format: "date" } },
  },
});

Where the standard stands

As of this scanner's build date, WebMCP is a W3C Community Group proposal. The only shipping implementation is an origin trial in Chrome (from Chrome 146 Canary, February 2026). Stable, cross-browser support is not expected until later in 2026 at the earliest. There is also no published, centralized discovery mechanism — an agent currently has to visit a page to find out whether it exposes tools.

Why this matters for scanning

Because tool registration happens in JavaScript at runtime, a scanner that only fetches and reads HTML (as MCPWebsite's V0.1 scanner does) cannot execute that code and confirm registration actually happened. That's why our reports show a detected / not detected / unable to verify status for WebMCP instead of a simple yes/no — see our State of WebMCP page for the detail.