WebMCP | AIEndpoint Docs

AIEndpoint and WebMCP are two halves of one loop. Your /.well-known/ai manifest answers "which site can do this?" before an agent navigates; WebMCP answers "what can I do on this page?" after it arrives.

AIEndpoint finds the right website. WebMCP operates it.

Overview

WebMCP is an emerging browser standard: a page registers structured tools via document.modelContext.registerTool(...), and an in-browser agent (Chrome 149+ behind a flag, or the ChatGPT desktop browser) can call them directly instead of guessing at your UI.

AIEndpoint connects to it in three places:

Your manifest can declare WebMCP support, so agents know your site is operable before they open it.
@aiendpoint/webmcp promotes your manifest's capabilities into live WebMCP tools - names, descriptions, and input schemas derived from what you already published.
The registry detects support at validation time, shows a ⚡ WebMCP badge, and lets agents filter for operable sites.

Declare support

Add one key to your manifest's meta object (the spec's extension point):

{
  "aiendpoint": "1.0",
  "service": { "name": "DemoWeather", "description": "..." },
  "capabilities": [ ... ],
  "meta": {
    "last_updated": "2026-09-01",
    "webmcp": "true"
  }
}

Agents reading your manifest see it immediately, and the registry validator records it. Even without the declaration, the validator also scans your homepage for the WebMCP API surface (modelContext, registerTool(...)) as a fallback heuristic.

The adapter

@aiendpoint/webmcp turns the capabilities you already declared into page tools. You supply one explicit handler per capability - nothing is exposed that you did not wire up.

<script src="/static/aiendpoint-webmcp.iife.js"></script>
<script>
  AIEndpointWebMCP.registerAiEndpointTools({
    specUrl: "/.well-known/ai",
    handlers: {
      current_weather: async (input) => {
        const data = await loadCity(input.city);
        updateWeatherCard(data); // drive the UI the human is watching
        return data;
      },
    },
  });
</script>

Or as an ES module in a bundled app:

import { registerAiEndpointTools } from "@aiendpoint/webmcp";

registerAiEndpointTools({ specUrl: "/.well-known/ai", handlers: { ... } });

What the adapter does for you:

Parses the spec's compact parameter notation ("string, required -- city name") into JSON Schema inputSchema with types and required lists
Probes every known API surface (document.modelContext, navigator.modelContext, legacy entry points), preferring registerTool with AbortSignal cleanup
Keeps a page-wide tool union on provideContext-only runtimes so independent components never clobber each other
Normalizes tool input (runtimes pass parsed objects or raw JSON strings)

See it live: weather.aiendpoint.dev - the badge in the corner reports registered tools, and agent calls update the weather card on screen.

The WebMCP badge

Services detected as WebMCP-capable get a ⚡ WebMCP badge on the services directory, next to the verified badge. Detection happens at registration and validation time:

meta.webmcp is "true" in the manifest (authoritative), or
the homepage HTML references the WebMCP API surface (heuristic)

For agents

The list API accepts webmcp=true:

curl "https://api.aiendpoint.dev/api/services?q=weather&webmcp=true"

And on aiendpoint.dev itself, the find_services WebMCP tool accepts webmcp_only: true - so an in-browser agent can ask the registry for sites it will be able to operate after navigating, closing the discovery-to- execution loop in one session.

Browser support

Chrome 149+ with chrome://flags/#enable-webmcp-testing enabled
ChatGPT desktop app built-in browser
Diagnostics: /webmcp/ping registers a single ping tool and reports exactly which API surface your browser exposes

The API surface is still evolving; the adapter is deliberately defensive and degrades to a no-op in unsupported browsers.