What is WebMCP?

01 How it works

Each action is called a tool. A tool has a name, a short description of what it does, and a list of the inputs it needs, written as a JSON Schema. The page hands its tools to the browser. When an AI agent is on the page, it reads the list, picks the tool that fits the task and calls it with real values. The site’s own code does the work and sends back the result.

When the agent runs in the person’s browser, it works inside their session, with the same login and cart they see. And because the site decides which tools exist, the site also decides what an agent can do there.

02 Two ways to add tools

From an HTML form

If your site already uses plain HTML forms, add a name and a description to the form. The browser turns it into a tool, and the labels you wrote for people describe the inputs to agents. Your server receives the same request it always did.

<form toolname="book_table"
      tooldescription="Book a table at the restaurant."
      action="/reservations">
  <label for="date">Date</label>
  <input type="date" id="date" name="date" required>

  <label for="guests">Guests</label>
  <input type="number" id="guests" name="guests" min="1" max="12" required>

  <button type="submit">Book</button>
</form>

From JavaScript

For anything a form can’t do, such as filtering a list that’s already on the page, register a function. You define the inputs and write the code that runs.

document.modelContext.registerTool({
  name: "search_products",
  description: "Search the catalog and return matching products with prices.",
  inputSchema: {
    type: "object",
    properties: {
      query: { type: "string", description: "What the shopper is looking for" },
      maxPrice: { type: "number", description: "Highest price, in US dollars" }
    },
    required: ["query"]
  },
  async execute({ query, maxPrice }) {
    const products = await searchCatalog(query, maxPrice);
    return { content: [{ type: "text", text: JSON.stringify(products) }] };
  }
});

Older guides use navigator.modelContext. The spec has moved the API to document.modelContext, and Chrome 150 and later deprecate the old name.

03 Where it works today

WebMCP is still an experiment, so support changes from month to month. ChatGPT is the first major assistant to use it. OpenAI’s docs say: “Site tools are ChatGPT's implementation of the proposed WebMCP standard.”

As of October 5, 2026. The ecosystem tracker has the full list, with a source for every entry.

04 Why websites add it

Agents that read the screen take a screenshot, guess where to click and check what happened, over and over. That is slow, and it breaks easily: a redesign or a cookie banner can stop an agent that worked yesterday. A tool call skips the guessing.

In the first run of our WindTunnel benchmark (August 2026), agents using WebMCP were up to 5.5× faster and 23× cheaper than agents working the screen, and solved 98% of the tasks. The live leaderboard has the latest results.

Tools also show you what agents are trying to do on your site, because every call names the action and its inputs. You choose which actions exist, and anything that spends money or makes a commitment should ask the person to confirm first.

05 WebMCP, MCP and llms.txt

MCP, the Model Context Protocol, connects an AI app to a server somewhere else, much like an API. WebMCP works in the browser: the page itself hands tools to the agent that is using it. llms.txt is a text file that helps AI read your site; WebMCP lets AI use it. They do different jobs, so a site can use more than one.

Chrome’s guide explains when to use WebMCP and when to use MCP.

07 Where to go next

• Add tools to your HTML forms, step by step →
• Build WebMCP tools for your website →
• Browse websites that already support WebMCP →

08 Sources

• WebMCP proposal and draft specification, W3C Web Machine Learning Community Group
• WebMCP documentation, Chrome for Developers
• Google I/O 2026 announcements, Google
• Site tools in ChatGPT, OpenAI
• WebMCP on Chrome Status
• WindTunnel benchmark and ecosystem tracker, webmcp.com