Skip to content
AI for SMBs

WebMCP: what it is, how it works and when it makes sense

Wouter·15 May 2026·7 min read
TL;DR

WebMCP is a proposal for a web standard that lets a website itself tell an AI agent in the browser which tasks can be performed, such as filling in a form or requesting an appointment. It is not a standard yet: Chrome offers it as a trial from version 149 and the specification is a draft. We have not built it into mindsora.io yet. For most business sites a well-built form comes first.

An AI agent that wants to book an appointment on someone's behalf has to guess today. It looks at your page, searches for something that resembles a form and tries to fill in the fields. WebMCP wants to make that guessing unnecessary: your website states what is possible. Below is what the proposal contains, how far along it is and what we did with it ourselves.

What is WebMCP?

WebMCP is a proposal for a web standard with which a website offers tasks to an AI agent in the browser. Chrome calls it in its documentation "a proposed web standard to help you build and expose structured tools for AI agents". Such a task is called a tool: a function with a name, a description and a fixed list of what the agent has to supply.

Chrome describes the difference with the usual way of working like this: the agent no longer has to study a button or a field to understand what it is for, because "the website declares the element's purpose".

Chrome announced WebMCP on 10 February 2026 as an early preview program for developers. The documentation overview page carries the publication date 18 May 2026.

Is WebMCP a standard already?

No. The specification is a draft report by the W3C's Web Machine Learning Community Group, with editors from Microsoft and Google. The document itself says: "It is not a W3C Standard nor is it on the W3C Standards Track."

In Chrome you can try it through an origin trial from version 149, or locally behind a flag. web.dev warns that WebMCP is "in active development" and may change. Chrome writes that the goal is for any browser with an agent to be able to implement it. Whether and when other browsers will do so is not in these sources.

How does WebMCP work?

In Chrome, WebMCP works in two ways: with a few extra attributes on an HTML form, or with JavaScript. The draft specification only describes the JavaScript route; the form attributes are in Chrome's documentation.

The declarative way. You give an existing form a name and a description. The browser turns it into a tool and uses the fields as input. When an agent calls the tool, the browser fills in the form while it stays visible to the user.

<form toolname="requestIntroduction"
      tooldescription="Requests an introductory call."
      action="/en/contact" method="post">
  <label for="name">Name</label>
  <input type="text" id="name" name="name">
  <label for="email">Email address</label>
  <input type="email" id="email" name="email">
  <button type="submit">Send</button>
</form>

If a field has no description of its own, the browser uses the text of the associated label. So good labels help here too. In this example the user submits the form: according to Chrome they have to click the button, unless you configure the form otherwise.

The imperative way. For tasks that do not fit in a form, you register a tool with JavaScript. That is work for a developer.

The example above is an illustration and is not on our site.

What is the difference between WebMCP and MCP?

MCP connects an AI application to your systems, wherever that application runs. WebMCP only works in the browser, while the user is on your page. The Model Context Protocol describes itself as "an open-source standard for connecting AI applications to external systems".

Chrome uses a comparison: MCP is a call center that is available anywhere and at any time, WebMCP is the in-store expert, who is only there while you are inside. Chrome also writes: "WebMCP is not an extension or a replacement of MCP."

How AI-ready is your business?

Take the free AI Readiness Quiz: 7 questions, 2 minutes.

Take the quiz →

What are the limits of WebMCP?

Chrome names three itself:

  • Discoverability. An agent only knows your site has tools once it visits the site. So WebMCP does not make an agent find you.
  • Meant for a browser with a person present. Chrome says it is primarily designed for a local browser with a human in the loop; headless use may be possible.
  • Extra work for complex sites. A site with many states will probably need extra code.

Security comes on top of that. In its security guidance Chrome warns that language models are susceptible to hidden instructions in text (prompt injection) and that safety inside a language model cannot be guaranteed. For an action with consequences, such as a booking or a payment, you can mark a tool registered with JavaScript so that the browser or the agent can ask the user for confirmation first.

Should you build in WebMCP now?

For most business sites, not yet. It is in a trial phase, the interface may change and it only helps an agent after it has already found your site. Google writes in its guide to generative AI in Search that protocols such as the Universal Commerce Protocol are "emerging". Google gives its advice to look into agents under a condition: "relevant to your business and you have extra time". That guide does not mention WebMCP by name.

What you can do now costs little and stays useful:

  1. Give every form field a label. The declarative way of WebMCP reads those labels, and a screen reader already does so today.
  2. Use real buttons and links. web.dev advises that for agents as well.
  3. Keep your most important task simple. Requesting an appointment in one form seems easier to us, for a person and for an agent, than in five steps.

If you have a web shop or a booking system with a lot of traffic and a developer at hand, Chrome's trial is worth trying. Chrome names customer support, travel booking and shopping as examples.

More about the basics is in what an agent-ready website is and in our checklist.

Has mindsora.io built in WebMCP?

No. What we did and did not do:

  • Not built in. In our Lighthouse measurement on 13 November 2026 the WebMCP checks did not apply. On 14 November we passed 3 of the 3 checks that did apply (own measurement through PageSpeed Insights, homepage, mobile).
  • Why not yet. It is a trial whose interface may change. Whether and when we build it in has not been decided.
  • What is there. A service catalogue as JSON, so an agent can fetch our packages and add-on services with prices without parsing the page.

What we do not know: whether there are agents visiting our site and whether WebMCP would change anything about that.

Frequently asked questions

Do I have to rebuild my website for WebMCP? Not for an ordinary form. The declarative way adds attributes to an existing form. For a highly complex site Chrome says you likely need to refactor or add JavaScript.

Does WebMCP replace an API? No. An API is for systems that talk to each other directly. WebMCP is for an agent in your visitor's browser.

Can an agent just buy something through WebMCP? You can mark a consequential tool registered with JavaScript as such, so the browser or the agent can ask for confirmation first. With a form like the one in the example, the user clicks submit. Chrome is frank that safety in a language model cannot be guaranteed.

Does WebMCP work in every browser? These sources only mention Chrome, as a trial from version 149.

Our free Agent-Ready Grader tests other points, among them crawler access, llms.txt, schema.org and page structure (header, main, nav), with a score from 0 to 100. It does not test buttons, labels, layout stability or WebMCP.

Want to work out how many hours a recurring task costs you? The free AI Readiness Scan does that: 7 questions about your business processes.

Ready to automate?

Book a free consultation and discover what AI automation can do for your business. Or take the 2-minute AI Readiness Quiz first.

Book a call

More articles