WebMCP is a proposed web standard that lets a website hand AI agents its forms and functions as named tools. An agent working in a buyer’s browser can then fill in a demo request or run a pricing calculator by calling the tool, instead of guessing which box does what.
It is still an experiment, running in Chrome and Edge. For a B2B company it comes after the basics of AI search and digital visibility: being found and read by AI tools is about pages, while WebMCP is about what an agent can do once it arrives.
How WebMCP works
Today a browser agent uses a site the way a person would, which Chrome’s documentation calls actuation: it reads the page, then clicks and types. Every step is a guess about what a button or a field is for, and each guess is a chance to get it wrong.
WebMCP replaces the guesses with declarations. The site registers tools, each with a name, a plain description and a schema of the inputs it takes, and the agent calls them. Chrome’s WebMCP documentation describes two ways to build one:
- The declarative API turns an existing HTML form into a tool with a few attributes on the form and its fields. Where a field has no description of its own, the browser uses its label.
- The imperative API registers JavaScript functions through
document.modelContext, for anything a form cannot express, such as filtering results or changing a setting.
Tools run inside the open tab, with the visitor’s own session, and disappear when the tab closes. When an agent calls a form tool, the browser brings the form into view and fills it in, and unless the site lets the agent submit, the person presses the button.
WebMCP vs MCP, llms.txt and schema markup
The names invite confusion. Chrome’s team says plainly that WebMCP “is not an extension or a replacement of MCP”. MCP connects an agent to a company’s servers from anywhere, at any time. WebMCP works only on the page the person has open, through their browser.
| Where it lives, and for whom | What it gives an agent | |
|---|---|---|
| WebMCP | In the page, for agents in the visitor’s browser | Actions: fill in a form, run a search, book a slot |
| MCP | On a company server, for any client connected to it | Data and background tasks, at any time |
| llms.txt | A file at the site’s root, for tools that choose to read it | A list of pages worth reading |
| Schema markup | In the page, for search engines and crawlers | Facts about what the page describes |
Only WebMCP lets an agent act. The other three help it find and read, and an llms.txt file in particular is read by fewer systems than its reputation suggests.
Where WebMCP stands today
WebMCP arrived in Chrome 146 behind a testing flag. It opened as an origin trial in Chrome 149 in June 2026, which lets a registered site switch it on for real visitors. In late September Google extended the trial through Chrome 162, saying feedback had “suggested improvements to the API”.
Support stops there. Chrome Status lists no signal from Firefox or from WebKit, the engine behind Safari. Microsoft runs its own trial in Edge, open until 30 March 2027, and engineers at Google and Microsoft edit the spec.
That spec is a draft from the W3C’s Web Machine Learning Community Group, which says it is not a W3C Standard nor on the Standards Track. Chrome’s documentation calls WebMCP “under active discussion and subject to change in the future”.
The agents are coming later than the API. At Google I/O in May, Google said Gemini in Chrome “will soon support WebMCP APIs”. For testing, Chrome’s documentation points developers to two extensions that read and call a page’s tools.
Registering for the trial gives a token for a meta tag or a response header. Without one, a site’s tools register only in browsers with the testing flag switched on, which in practice means its own developers.
Does WebMCP help SEO or AI citations?
Not directly. Chrome lists as a limitation that clients and browsers “must visit a site directly to know if it has callable tools”, and nothing in its documentation ties WebMCP to search. Being cited still depends on the passage an answer can lift.
Where it counts is the step after the answer. A buyer who asks an assistant to shortlist vendors can then ask it to book a call with one, and on that step a form the agent can call is more reliable than one it has to work out from the screen.
Chrome already measures it. Lighthouse’s Agentic Browsing category, announced in June, checks a page for registered tools, for forms missing declarative WebMCP and for valid schemas. It reports checks passed, not a score out of 100.
What a B2B site would expose
Chrome’s examples lean on shopping and travel, but the first use case in its overview is a support flow for a software company. A B2B site has a short list of tasks a buyer’s agent would be sent to do:
- Request a demo or a quote, through the form the site already has.
- Book a meeting, where a date picker built for people is hard for an agent to read.
- Check pricing or run a calculator, with the inputs a buyer would type.
- Search the documentation or open a support ticket, for customers already signed in.
Chrome’s best practices say each tool should do a single job, and a tool should be registered only while it is useful in the current state of the page. A tool for every link on the site defeats the point.
On a demo request form, the declarative version is a handful of attributes:
<form toolname="request_demo"
tooldescription="Ask for a product demo. The sales team replies by email.">
<label for="company">Company</label>
<input id="company" name="company" required>
<label for="team">Team size</label>
<input id="team" name="team" type="number"
toolparamdescription="How many people would use the product">
<button type="submit">Request a demo</button>
</form>
Leave out toolautosubmit, the attribute that lets an agent send the form itself. The buyer then reviews what the agent filled in and presses the button, so every lead that arrives is one a person approved.
One catch is common in B2B. A form embedded from a marketing platform usually sits in a cross-origin iframe, where tool registration is off by default. The page has to add allow="tools" to the iframe, and the form inside it has to carry the attributes, which is the platform’s code.
How to tell which leads an agent filled in
WebMCP tells the page when an agent acts. Once an agent has filled in a form tool, the browser fires a toolactivated event that names the tool, and a toolcancel event if the person stops it.
document.modelContext?.addEventListener('toolactivated', ({ toolName }) => {
if (toolName === 'request_demo') {
document.querySelector('[name="filled_by"]').value = 'agent';
}
});
The snippet writes to a hidden field on the form. Store that field on the contact record in the CRM, next to the first AI touch and the buyer’s own answer about how they found you. When the agent also submits the form, its submit event carries agentInvoked set to true.
The field earns its place in the pipeline report. Agent-filled leads may close better or worse than the rest, and only reporting that follows a lead to revenue will show which.
The security side of WebMCP
A tool is a door, and agents can be tricked into using it. Chrome’s security guidance warns that models are open to indirect prompt injection, and that “it’s impossible to guarantee safety inside of a large language model”. Its advice for tool builders:
- Mark a tool that returns content written by others, such as reviews or forum posts, with
untrustedContentHint. - Mark a tool that does something significant with
consequentialHint, so the agent or browser can ask the person first, and a tool that changes nothing withreadOnlyHint. - Keep tools to your own site. Other sites and cross-origin iframes cannot see them unless you name those origins in
exposedTo. - Keep the text short. Chrome suggests up to 500 characters per tool description, 150 per parameter description and 1,500 per tool output.
Should you add WebMCP now?
For most B2B sites, not first. The work that pays today is the work agents and people share: pages that state pricing, integrations and proof, and forms with real labels, which the declarative API reuses as descriptions anyway.
After that, WebMCP is a small and reversible test. Annotate the demo form, register for the origin trial and check the page in Lighthouse. Chrome describes WebMCP as a progressive enhancement, so the form keeps working in every browser that ignores the attributes.
Wait if the forms that matter sit inside an embed you cannot edit, or if the only tools worth having would act on customer accounts. Those need the security work above before they are worth a trial.
● About the author
Hamza Hashim is the marketing strategist at Katama. He owns the channel mix and the order it gets built in: what ships first, what waits for evidence, and what gets stopped when the evidence says stop. Meet the team.