WebMCP vs MCP: what it is and how to check your site

• Check a website
• What WebMCP is
• WebMCP vs MCP
• Declarative example
• Imperative example
• What the check sees
• Test in a real browser
• Sources

What WebMCP is

An AI agent that works inside a browser normally has to read the page and guess: which box is the search field, which button sends it? WebMCP removes the guessing. The page publishes tools, small functions with a name, a plain-language description and a list of inputs. The agent calls the tool instead of clicking around, and the person can watch the page react.

A page can offer tools in two ways:

• Imperative (JavaScript). The page calls document.modelContext.registerTool() with a name, a description, an inputSchema (a JSON Schema that describes the inputs) and an async execute function that does the work.
• Declarative (HTML). A normal <form> becomes a tool when you add toolname and tooldescription to it and toolparamdescription to its fields. The optional toolautosubmit attribute lets the browser submit the form when an agent calls the tool. Without toolparamdescription, the browser falls back to the field’s label text.

Status, as we read it on 2026-10-07. The specification is a Draft Community Group Report of the Web Machine Learning Community Group (the version we read is dated 2 October 2026), not a finished standard. Chrome’s documentation says the API is in an origin trial from Chrome 149, and for local tests names the flag chrome://flags/#enable-webmcp-testing. The declarative part is still a to-do section in the draft; its details live in the Declarative API explainer and in Chrome’s documentation. Names and behaviour can change. We do not promise support in every browser, a listing anywhere or a better search ranking.

WebMCP vs MCP

Both give AI agents tools with names, descriptions and inputs. They live in different places. The table follows Chrome’s own comparison.

Which one do you need? If assistants should call your data or actions from anywhere, that is MCP: see our own remote MCP endpoint as an example. If you want an agent in the browser to use your live page, that is WebMCP. Chrome’s documentation suggests combining them: core logic in MCP, a contextual page UI in WebMCP. Many websites need neither. A page with real text in its HTML is already readable for agents.

Small example: a form as a tool (declarative)

<form toolname="search_products"
      tooldescription="Search the product catalogue by keyword."
      action="/search" method="get">
  <label for="q">Search words</label>
  <input id="q" name="q" type="text"
         toolparamdescription="Search words, for example red shoes.">
  <button type="submit">Search</button>
</form>

Small example: a function as a tool (imperative)

// Inside an async function or a module script.
// Only some browsers have WebMCP today, so detect it first.
if (document.modelContext && typeof document.modelContext.registerTool === 'function') {
  await document.modelContext.registerTool({
    name: 'search_products',
    description: 'Search the product catalogue by keyword. Returns up to 10 products.',
    inputSchema: {
      type: 'object',
      properties: { query: { type: 'string', description: 'Search words, for example red shoes.' } },
      required: ['query']
    },
    execute: async ({ query }) => {
      const answer = await fetch('/api/search?q=' + encodeURIComponent(query));
      return await answer.text();
    }
  });
}

What our check sees, and what it cannot

The form above runs the WebMCP part of our agent readability check (check D3). It reads the homepage HTML (the first 1 MiB), within the same ten-second, nine-file budget as every agent check. No script of the website runs and nothing is stored.

In our score, WebMCP is worth 8 of the 100 points (group D, with llms.txt and MCP discovery). The CheckWebsiteNow Agent Readability Score is our own transparent score, not an official standard, and there is no official WebMCP score. Every check counts with its weight, and a part we could not measure never counts as passed. How the score works. A busy or limit message comes from us and says nothing about the website: we allow ten agent checks a minute per client.

How to test WebMCP in a real browser

1. Use a browser that has WebMCP switched on. In Chrome that means the origin trial or the local flag named in Chrome’s documentation. Other browsers may not have it yet.
2. List what the page registered. The draft describes a getTools() method on document.modelContext for discovery, and Chrome’s documentation lists it too. Call it from the developer console on your own page.
3. Run Lighthouse. Lighthouse 13.5 (released 18 September 2026) added WebMCP audits such as webmcp-registered-tools and webmcp-form-coverage. They run in a real browser. Ours does not, so use both when WebMCP matters to you.

Sources

• WebMCP specification draft (Draft Community Group Report, 2 October 2026; read on 2026-10-07).
• WebMCP explainer on GitHub, Web Machine Learning Community Group.
• Chrome documentation: WebMCP, with the Imperative API, the Declarative API and the page WebMCP versus MCP (read on 2026-10-07).
• Lighthouse 13.5.0 release notes (18 September 2026).
• On this site: agents, MCP and the score, the API guide and the OpenAPI description of POST /api/v1/agent-check.