Implementation guide · kept current with the WebMCP spec · updated 29 Aug 2026

Make your site agent-callable

“Callable” means an AI assistant can do something on your site — not just read it. Getting there is one script tag. This page is the current, correct way, kept up to date as the standard changes (it renamed itself in May; it will change again).

Before you start: is this the right rung?

Climb in order. If your scan says agents can't read you (rung 0) or can't answer buyers' questions about you (rung 1), fix those first — a text file and some published answers, no code at all. Tools amplify a legible site; they can't rescue an illegible one.

What WebMCP is, in one paragraph

A browser standard that lets your page hand AI assistants typed tools instead of buttons to guess at. As of 25 August 2026, ChatGPT's desktop-app browser executes these tools; Chrome is trialling the same standard. The code runs only in your visitor's tab, only when an agent-capable browser is present, and does nothing at all anywhere else.

The starter tool: your enquiry form, callable

Copy this, adjust the three selectors to match your form, and paste it into your site's custom-code box (Wix: Settings → Custom Code · Squarespace: Code Injection · Shopify: theme.liquid · WordPress: a header-scripts plugin — always the “before </body>” slot).

<script>
// Agent Surface Scan — starter WebMCP tool (API current as of late Aug 2026).
// Wraps your existing enquiry form as a tool an AI assistant can call,
// with the visitor confirming before anything is sent.
(function () {
  // Publish a detectable manifest FIRST — ordinary browsers don't implement
  // modelContext, so this is how scanners (and the Agent Surface Scan) can
  // verify your site declares tools:
  window.__webmcpToolManifest = ["send_enquiry"];
  document.documentElement.dataset.webmcpTools = "send_enquiry";

  if (!("modelContext" in document)) return; // harmless everywhere else

  document.modelContext.registerTool({
    name: "send_enquiry",
    description:
      "Send an enquiry to this business. Use when the visitor wants to " +
      "get in touch, ask about services, or request a callback.",
    inputSchema: {
      type: "object",
      properties: {
        name:    { type: "string", description: "Visitor's name" },
        email:   { type: "string", description: "Visitor's email address" },
        message: { type: "string", description: "What they need, in their words" }
      },
      required: ["name", "email", "message"]
    },
    async execute({ name, email, message }) {
      // 1. Fill YOUR existing form's fields (adjust the selectors):
      const form = document.querySelector("form");
      if (!form) return { content: [{ type: "text", text: "No form found on this page." }] };
      form.querySelector('[name=name]')?.setAttribute("value", name);
      form.querySelector('[name=email]')?.setAttribute("value", email);
      const msg = form.querySelector("textarea");
      if (msg) msg.value = message;

      // 2. Ask the HUMAN to confirm — write tools must never fire silently:
      const okay = confirm("Your assistant wants to send this enquiry:\n\n" + message + "\n\nSend it?");
      if (!okay) return { content: [{ type: "text", text: "The visitor declined to send the enquiry." }] };

      form.requestSubmit();
      return { content: [{ type: "text", text: "Enquiry submitted. The business will reply to " + email + "." }] };
    }
  });
})();
</script>

Note the confirm() step: a tool that acts on someone's behalf asks the human before it commits. That is the difference between Transactable and creepy.

Test it in two minutes

Open your site in the ChatGPT desktop app's browser (latest version).
Ask: “What tools does this page offer? Send an enquiry for me — my name is…”
Then re-scan your site here — the scanner detects registered tools, and your result page updates to show them.

The current API contract (for you or your AI)

If you're asking an AI assistant to draft tools for you, give it these facts — training data is usually behind this spec:

The entry point is document.modelContext — the older navigator.modelContext is deprecated; feature-detect with "modelContext" in document.
registerTool({ name, description, inputSchema, execute }) — inputSchema is JSON Schema; execute(args) receives an object (not a JSON string) and returns { content: [{ type: "text", text: "…" }] }.
Tools are gated by a tools permissions policy defaulting to same-origin; cross-origin iframes need allow="tools".
The tool description is what the model reads to decide when to call it — write it as instructions, and treat it as a security surface: a misleading description misleads the agent.
ChatGPT's implementation covers a subset of the spec (no declarative form attributes, no iframe tools) — keep to registerTool.

Want the full picture — which of your capabilities are worth exposing, in what order, at what value? That's the Agent Opportunity Map. This page gets you your first tool; the map gets you the right ten.