How AI agents use your website with WebMCP

What is WebMCP?

Browser agents such as Gemini in Chrome, Claude in Chrome or Perplexity Comet mostly operate websites the way a human does. They read screenshots or the DOM, look for the right field and click. That is slow and breaks as soon as the layout changes. With WebMCP, the website itself describes which actions exist, which inputs they need and what comes back.

The proposal was developed by teams at Google and Microsoft. The first explainer appeared on 13 August 2025. The specification is a Draft Community Group Report of the W3C Web Machine Learning Community Group. This means WebMCP is not yet an official W3C standard, and details may change.

WebMCP and MCP

The name is based on the Model Context Protocol, which Anthropic introduced in 2024 and handed over to the Linux Foundation's Agentic AI Foundation in December 2025. Both use the same idea of tools with a name, a description and an input schema. However, they work in different places.

How WebMCP works

Chrome offers two ways. The declarative variant extends existing HTML forms, the imperative one registers tools via JavaScript.

Declarative with forms

A form becomes a tool as soon as it carries toolname and tooldescription. The browser derives the input schema from the fields, their types and attributes such as required. With toolparamdescription you explain individual fields. Without toolautosubmit, the agent only fills in the form and the human submits it.

In the submit event, agentInvoked shows whether an agent triggered the form. Via respondWith() you return a structured result to it. According to the explainer, the declarative variant is not yet part of the specification. So the attribute names may still change.

Imperative via JavaScript

For configurators, carts or multi-step flows, you register a tool with document.modelContext.registerTool(). It needs a name, a description, an input schema as JSON Schema and a function that calls your existing logic.

With annotations you mark tools that only read (readOnlyHint) or have consequences (consequentialHint). Older guides use navigator.modelContext. The specification and the Chrome documentation now use document.modelContext.

What can browsers and agents do today?

By default, WebMCP tools are only allowed on the page itself and in frames of the same origin. Frames of a different origin need allow="tools". Chrome describes the interface as a tool for flows in the browser where a human is present.

Measuring what agents do

Browser agents that work via screenshot and click often look like fast human visits in log files and web analytics. WebMCP, on the other hand, delivers clear signals.

1. For imperative tools, you write an event directly in execute, as in the example above. The code runs on every call.
2. For forms, you evaluate agentInvoked in the submit handler and send it as a parameter to your web analytics.
3. Route the events into GA4, the dataLayer or your data warehouse so you can see the trend over weeks.
4. Check the registration in the DevTools panel and test individual tools with the Model Context Tool Inspector extension.

Common mistakes

• toolautosubmit is also set on forms that order, book or send data. Only let read-only actions such as a search submit automatically.
• Descriptions are vague (“Form 1”). The agent chooses tools based on the description.
• The code only uses navigator.modelContext and does not follow the current state of the specification.
• Tools sit in a frame of a different origin without allow="tools".
• Tools stay registered even though the action is currently not possible, for example with an empty cart.
• Someone expects better rankings or more mentions in ChatGPT. No provider has announced any such effect. WebMCP makes your website usable for agents.

Frequently asked questions