WebMCP: the business briefing

What it is

In plain language: WebMCP lets a website hand AI agents a short, labelled list of things they can do on it, such as “book an appointment” or “check stock”, with clear inputs. The agent calls those actions directly instead of guessing which button to click.

Technically: websites expose JavaScript functions or HTML forms as “tools”, each with a natural-language description and a structured input schema. AI agents in the browser, in iframes or in extensions can discover and call those tools.

Who is building it

WebMCP is being developed in the W3C Web Machine Learning Community Group. Engineers from Google and Microsoft wrote the specification together. It is a proposed web standard: a Community Group draft, not a finished W3C Recommendation.

Two ways to add it

1. Declarative: turn an existing form into a tool

Add attributes to an HTML form you already have. toolname and tooldescription go on the form, and toolparamdescription goes on each field.

<form toolname="request_quote"
      tooldescription="Request a delivery quote between two cities">
  <input name="from" toolparamdescription="City the goods leave from">
  <input name="to" toolparamdescription="City the goods go to">
  <input name="pallets" type="number" toolparamdescription="Number of pallets">
  <button type="submit">Request quote</button>
</form>

2. Imperative: register a tool in JavaScript

For richer tasks, register a tool with document.modelContext.registerTool(). Tools can be removed again with an AbortSignal.

if (document.modelContext) {
  await document.modelContext.registerTool({
    name: "check_branch_stock",
    description: "Check whether a product is in stock at a branch",
    inputSchema: {
      type: "object",
      properties: { sku: { type: "string" }, branch: { type: "string" } },
      required: ["sku", "branch"]
    },
    async execute({ sku, branch }) {
      // Run your own site code here and return the result to the agent.
    }
  });
}

In August 2026 the API moved from navigator.modelContext to document.modelContext. The old name is kept as an alias.

How it stays safe

• A person is present. WebMCP is designed around a human in the browser with the agent.
• Forms need a human to submit. Form tools wait for a person to press submit, unless the site owner allows auto-submit.
• Agent actions are visible. Sites can see when an agent triggered an action and can show visual feedback while it works.
• No passwords for the AI. Tools run in the website’s own code, using the visitor’s existing login.
• Nothing hidden is exposed. Site owners choose which tools exist. Nothing is available that the site does not register.
• Embedded content can be switched off. The Permissions-Policy header tools=() turns tools off for embedded content.

WebMCP and MCP are related but different. MCP, introduced by Anthropic in 2024, connects AI apps such as Claude and ChatGPT to a business’s backend server. WebMCP works in the visitor’s browser, on your public website, for AI agents working on behalf of a person.

What to do in the next 90 days

1. List your top customer journeys. Enquiries, bookings, quotes, stock questions and support.
2. Pick one or two low-risk actions first. Read-only information or an enquiry form is a safe place to start.
3. Add declarative attributes to an existing form. It is the quickest way to become agent-ready, and it keeps a person in charge of the submit button.
4. Decide your permission levels. Agree which actions may never happen without a person confirming.
5. Test, then expand. Test in the browsers that support WebMCP today, watch how agents use your site and add tools as the standard matures.