Make your page agent-ready with WebMCP

The shape

An AI agent lands on your page and wants to do something real — search your catalog, add to a cart, start a return. Today it guesses: it reads the DOM, finds something that looks like a button, and clicks it. That breaks the moment you ship a redesign.

WebMCP is the W3C answer. Your page declares what it can do, as typed tools with JSON Schemas, and the agent calls those instead of hunting for selectors.

The cheapest possible start

If a form already does the job, attributes are enough — no JavaScript:

<form toolname="search-products"
      tooldescription="Search the catalog by keyword and category"
      toolautosubmit>
  <input name="q" toolparamdescription="What to search for" required>
  <select name="category" toolparamdescription="Which department">
    <option value="all">Everything</option>
    <option value="shoes">Shoes</option>
  </select>
  <button type="submit">Search</button>
</form>

import { installDeclarativeTools } from "@dornick/webmcp/declarative";
installDeclarativeTools();

Hand the agent a real answer rather than a page reload:

form.addEventListener("submit", (e) => {
  e.preventDefault();
  const results = search(new FormData(e.target));
  render(results);
  if (e.agentInvoked) e.respondWith(Promise.resolve({ count: results.length, results }));
});

agentInvoked is true only when an agent pressed the button, so your human path is untouched.

Where Dornick changes the maths

Everything above is standard WebMCP — any toolkit gets you there.

The difference is that Dornick already knows your data points, so the provider surface comes free:

new Dornick({ config, transport, webmcp: { publish: true } });

// Registers dornick_set_email, dornick_set_plan, … one per configured dataPoint,
// each schema derived from the type / pattern / allowedValues you already wrote,
// plus the read-only workspace verbs and dornick_ask.

And it works in the other direction at the same time: Dornick’s own agent adopts every tool on document.modelContext — yours, and anyone else’s — as real callable tools named webmcp__*. One flag, both sides.

What to expect in each browser

Environment	Who can see your tools
Chrome 149+ with the origin trial or chrome://flags/#enable-webmcp-testing	the browser’s own agent, extensions, DevTools, page script
Anything else (polyfilled)	extensions, CDP/Playwright drivers, the MCP relay, page script

A polyfill cannot reach the browser’s built-in agent — that’s a browser IPC surface, not page JavaScript. Plan for native if browser-agent reach is the goal; ship the polyfill either way, because it costs ~7 KB and covers everything else.

FAQ