Skip to content
Technology

What is WebMCP – and why AI agents will soon be filling in forms on websites

WebMCP is an experimental web standard that lets a website tell AI agents in the browser which tools it offers – such as “request a project” or “estimate costs” – instead of the agent having to guess at buttons and fields. Chrome has been testing it in an origin trial since version 149; no ranking signal of its own is known. I have implemented four tools on my own website, and I offer the same for existing forms at my hourly rate of €95. This guide explains how WebMCP works, what it delivers today and what I learned while implementing it (as of 28 October 2026).

Published
Last updated:
Reading time
8 min read
Author
By Maurice de Felice, web developer in Augsburg, Germany

What problem does WebMCP solve?

AI agents that act on a user’s behalf in the browser – booking an appointment, requesting a quote, comparing prices – today mostly work like someone wearing a blindfold: they read the page structure, guess which field is which, and click. If the layout changes, the flow breaks.

WebMCP turns this around. The website explains itself: “I have a tool called projekt_anfragen with the fields name, email and message.” The agent receives a machine-readable schema and no longer has to guess anything. The name borrows from the Model Context Protocol (MCP), which connects AI assistants to tools – but here everything runs directly in the browser, on the page the user currently has open.

How does WebMCP work technically?

A declarative form tool can also carry the toolautosubmit attribute; the agent then submits the form itself. Without this attribute it only fills the form in, and the human decides whether to click “Send”. For contact forms, I think that is the right line to draw.

Source: Chrome documentation on WebMCP (declarative and imperative API), as of October 2026
ApproachHowWhat for
DeclarativeAttributes toolname and tooldescription on the <form>, toolparamdescription on fieldsExisting forms: contact, search, enquiry
Imperativedocument.modelContext.registerTool() with name, description, JSON schema and functionFunctions without a form, e.g. a calculation
UnregisteringPass an AbortSignal when registering, call abort() laterTools that only apply temporarily
Responding to the agentOn submit, check event.agentInvoked and call event.respondWith()Report a result back, e.g. “Enquiry sent”

Source: Chrome documentation on WebMCP (declarative and imperative API), as of October 2026

Where does WebMCP stand today?

  • Chrome has been testing WebMCP in an origin trial since version 149: websites register and enable the feature for their visitors with a token. According to the announcement in the Chromium project, the trial was originally planned to run until Chrome 156; in September 2026 an extension to Chrome 162 was requested. The binding current end date is on the origin trial page.
  • For your own development, WebMCP can be switched on in Chrome via chrome://flags/#enable-webmcp-testing.
  • The specification is a draft by the W3C Web Machine Learning Community Group – not an official W3C standard. Early versions used navigator.modelContext; the current Chrome documentation describes document.modelContext.
  • Other browsers have not committed yet. Without support, simply nothing happens on the website – which makes WebMCP well suited to progressive enhancement.

Does WebMCP bring more visibility on Google or ChatGPT?

The honest answer: no ranking effect is known, and I would not promise one to any client. WebMCP concerns agents that have already opened a page – not whether the page is found in the first place. For visibility on Google and in AI search, content, structure, structured data and loading speed still count.

What is already changing: in its “Agentic Browsing” category, Lighthouse checks whether forms are annotated, whether tools are registered and whether their schemas are valid. If you annotate properly early on, you are prepared for when agents in the browser become commonplace – and you have an argument for clients who ask whether their site is “AI-ready”.

How did I implement WebMCP on my own website?

On mauricedefelice.dev, four tools are active as soon as the browser supports WebMCP – three forms and one tool in JavaScript: the contact form is described as projekt_anfragen, the website check as website_check_vorbereiten and the cost calculator as kostenrechner_ausfuellen. The fourth is the imperative tool kosten_schaetzen, which uses the same pricing logic as the calculator and returns a price range with timeline and cost drivers as its response.

Without WebMCP, nothing changes for visitors: registration first checks whether document.modelContext exists and only then loads the tool module. There are no error messages in the console and no extra JavaScript for browsers without support.

  • Radio buttons: Chrome reads the description of a radio group from toolparamdescription on the surrounding <fieldset>, not on the individual options. Set on the options, it makes Lighthouse report “Add a description”.
  • Spam protection: an invisible honeypot field in the form would become part of the tool schema – and an agent might fill it in. That is why I moved it out of the <form>.
  • Reading values: an agent fills fields directly in the page structure. On submit, I therefore read the values from the form itself and not just from the React component’s state.
  • No automatic submission: the contact and website check forms deliberately do not carry toolautosubmit. The agent fills in, the human sends.
  • Embedded content: in the embeddable cost calculator on other people’s websites, I do not register any tools.

What did the Lighthouse measurement show?

Besides WebMCP, the category also counts whether the accessibility tree is clean, whether there is an llms.txt and whether the layout shifts. Before, the schema audit was not applicable because there were no tools; now it is applicable and passed.

Lighthouse 13.5, mobile, Chrome with WebMCP switched on, measured on 27 October 2026; before = live website without WebMCP
PageAgentic Browsing beforeAgentic Browsing after
Home page5 of 5 audits (WebMCP schemas not testable)6 of 6 audits
Web app development5 of 5 audits (WebMCP schemas not testable)6 of 6 audits
Cost calculatorPage did not exist yet6 of 6 audits

Lighthouse 13.5, mobile, Chrome with WebMCP switched on, measured on 27 October 2026; before = live website without WebMCP

What should you do for your website now?

  • Build forms properly: real <label> elements, meaningful name attributes, mandatory fields marked as required. That helps agents, screen readers and your visitors alike – and it is the foundation for WebMCP.
  • Describe important forms declaratively: contact, quote, appointment request. A few attributes are all it takes.
  • Leave submission to the human whenever enquiries, orders or money are involved.
  • Do not promise anything: WebMCP is experimental and, as things stand, does not affect rankings. It is worthwhile as preparation, not as an SEO measure.
  • If you would like it implemented: for existing forms I charge by effort at my hourly rate of €95; for new websites and web apps it can be part of the fixed price on request.

What do clients often ask?

An experimental web standard that lets a website offer its functions to AI agents in the browser as tools with a description and a parameter schema.

No. Chrome has been testing WebMCP in an origin trial since version 149, and there is a flag for developers. Other browsers have not committed yet. Without support, the website stays exactly as it is.

Only if the form carries the toolautosubmit attribute. Without it, the agent fills the form in and the human decides whether to send it – which is how I implemented it on my website.

No such effect is known. WebMCP concerns agents on a page that is already open, not whether the page can be found.

Planning a project?

In a free initial call you get an honest assessment of scope, timeline and a fixed price – by video, or in person in Augsburg and Munich. I usually reply within 24 hours.

Maurice de Felice, web developer in Augsburg, Germany

About the author

Maurice de Felice

Freelance web developer in Augsburg, Germany. By day a backend developer at one of Germany’s largest comparison portals; as a freelancer I build websites, web apps and SaaS platforms with Laravel and React – for example the SaaS platform YachtBrokerBid. More about me

More guides: How much does a website cost in 2026? Prices, examples and cost drivers · Getting a website built: freelancer, agency or website builder? · How long does it take to have a website built?

Contact

Tell me what you need

Tell me what you are planning – a new website, a web app or changes to an existing one. I usually reply within 24 hours with a first assessment and a suggested time for a free initial call; after that you receive a fixed-price offer. I am based in Germany and work in English or German.

Send a message

What should be built, by when, and is there an existing website? Plus a quick note on how you found me.

I only use your details to answer your enquiry – you can request information or deletion at any time. Details in the privacy policy (German).