Today's AI agents use websites the hard way. They take screenshots, read the page's HTML, guess which button does what, and click or type like a slow, nervous human. When a layout changes or a modal pops up, the agent gets lost. WebMCP is a proposal to fix that by letting a website tell an agent, in a structured way, exactly what actions it offers and how to call them.
WebMCP has moved quickly from idea to browser experiment. As of October 2026 it is a draft specification being developed in the open by engineers from Google and Microsoft, with an origin trial available in Chrome. It is not finished, it is not supported by every browser, and no mainstream agent depends on it yet. But it is concrete enough that marketers and developers should understand what it is and what to prepare.
This guide explains what WebMCP is, how its two APIs work, where it actually stands today, and the low-risk steps you can take now to make your site easier for AI agents to use.
Key Takeaways
- WebMCP lets a web page expose "tools" (structured actions such as search, book or submit) that a browser-based AI agent can call directly instead of simulating clicks.
- It has two approaches: a declarative API that annotates existing HTML forms, and an imperative API that registers JavaScript functions.
- It is still a draft. Chrome ran an early preview from February 2026 and announced an origin trial starting in Chrome 149 at Google I/O in May 2026.
- WebMCP does not affect Google rankings and is not an SEO signal; it is about agents completing tasks on your site.
- The safest preparation today is clean, well-labelled forms and semantic HTML, which help agents, accessibility and conversions whether or not WebMCP ships widely.
What is WebMCP?
WebMCP (Web Model Context Protocol) is a proposed browser API that lets a website publish a set of tools for AI agents. Each tool has a name, a plain-language description and a schema describing its inputs. An agent running in the browser can discover those tools and call them, and the site returns structured results.
The name nods to the Model Context Protocol (MCP), the open protocol many AI applications use to connect to external tools and data. The difference is where the tools live. Classic MCP servers run on a back end and usually need separate authentication. WebMCP tools run inside the web page, in the user's browser tab, and inherit the user's existing session. Semrush's March 2026 explainer notes that WebMCP currently covers tools only, not MCP's other building blocks such as resources and prompts.
Chrome's team summarises the goal simply: build structured tools for agents to complete tasks efficiently and accurately. They also stress that it is meant to be cross-browser.
“Our goal is to build APIs that any browser with agentic capabilities can implement and benefit from.”
— Chrome for Developers, Join the WebMCP origin trial
If you are new to the idea of AI systems acting on behalf of users, start with our explainer on agentic search and our broader guide on how to optimise your website for AI agents.
Where WebMCP stands in October 2026
Because WebMCP is moving fast, dates matter. Here is the timeline based on Chrome's own announcements and independent tracking.
| Date | Milestone | What it means |
|---|---|---|
| August 2025 | Proposal published by Microsoft Edge engineers, with Google joining as co-author | WebMCP becomes a joint cross-browser effort rather than a single vendor's idea |
| February 2026 | Chrome early preview programme opens; experimental support behind a flag in Chrome 146 | Developers can test locally, but nothing ships to normal users |
| 18 May 2026 | Chrome publishes WebMCP developer documentation | Declarative and imperative APIs are documented officially |
| 19 May 2026 | Origin trial announced at Google I/O, starting in Chrome 149 | Sites can register for a token and test with real Chrome users |
| Mid-2026 | Draft specification updated; API surface moving from navigator.modelContext to document.modelContext | Code written today may need updating as the draft changes |
The work is being incubated in the W3C Web Machine Learning group. Patrick Brosset of Microsoft wrote in February 2026 that it was "still in draft form" and that Microsoft and Google were working closely together on it. Independent tracking by Spronta in July 2026 reported experimental support in Microsoft Edge behind a flag, and no implementation commitment yet from Firefox or Safari. It also noted that no mainstream AI agent was calling WebMCP tools in production; Chrome's documentation lists Gemini in Chrome support as "coming soon" without a date.
In short: WebMCP is real, documented and testable, but it is an experiment, not a web standard you can rely on for revenue.
How WebMCP works: two APIs
The declarative API (annotated HTML forms)
The declarative API turns an existing HTML form into a tool by adding attributes. Chrome's documentation lists toolname and tooldescription on the <form>, an optional toolparamdescription on individual fields, and an optional toolautosubmit attribute. The browser reads the form's fields and generates a JSON input schema. A <select> becomes a list of allowed values, and labels can serve as parameter descriptions.
<form toolname="createSupportRequest"
tooldescription="Submits a request for customer support.">
<label for="email">Email</label>
<input id="email" name="email" type="email" required>
<label for="topic">Topic</label>
<select id="topic" name="topic">
<option>Billing</option>
<option>Technical</option>
</select>
<button type="submit">Send</button>
</form>When an agent calls the tool, the browser focuses the form and pre-fills it, and the form stays visible to the user. Without toolautosubmit, the user still clicks Submit. Chrome also adds an agentInvoked flag to the submit event so your analytics can tell agent-filled submissions from manual ones.
The imperative API (JavaScript tools)
The imperative API lets developers register tools in JavaScript, which suits actions that are not a single form, such as adding an item to a basket or filtering a product list. Each tool has a name, description, input schema and a function that runs when the agent calls it. Tools can be added or removed as the page state changes. The simplified sketch below shows the shape only; check Chrome's current documentation before building, because the draft is still changing.
// Illustrative only: the API location changed during 2026
const mc = document.modelContext ?? navigator.modelContext;
if (mc) {
mc.registerTool({
name: "searchProducts",
description: "Search the catalogue by keyword and maximum price.",
inputSchema: {
type: "object",
properties: {
query: { type: "string" },
maxPrice: { type: "number" }
},
required: ["query"]
},
async execute({ query, maxPrice }) {
const results = await runSiteSearch(query, maxPrice);
return { content: [{ type: "text", text: JSON.stringify(results) }] };
}
});
}Security and permissions
Both APIs are governed by a "tools" Permissions Policy that defaults to same-origin contexts, so a third-party iframe cannot expose tools unless you allow it. Tool calls need an open tab or webview; they do not run headlessly. That design keeps the human in the loop, but it also means you still need server-side validation, rate limits and confirmation steps for anything sensitive, such as payments.
WebMCP vs other ways of helping AI
| Approach | What it does | Status | Who it helps |
|---|---|---|---|
| WebMCP | Exposes actions on a page as callable tools for in-browser agents | Draft; Chrome origin trial | Agents completing tasks (booking, searching, submitting) |
| MCP server | Connects AI apps to your back-end tools or data | Widely used open protocol | AI apps and integrations, often with separate authentication |
| Structured data (schema.org) | Describes what a page is about: products, events, organisation | Mature, used by search engines | Search engines and AI understanding of content |
| llms.txt | Plain-text guide to a site's key content for language models | Informal proposal; not required by major engines | Some AI tools reading documentation |
These are complementary. Schema markup describes your content; WebMCP describes your actions. Our llms.txt guide explains why that file is optional.
How to prepare your site: a step-by-step plan
- List your key user actions. Quote requests, bookings, product search, store finder, account sign-up. These are the tools an agent would want.
- Clean up your forms. Every field needs a real
<label>, a sensiblename, the right input type and clear validation messages. Our guide to semantic HTML covers the basics. - Remove agent traps. Avoid CAPTCHA on low-risk forms where possible, unclear multi-step wizards and buttons that are really unlabelled icons.
- Track form completions properly. Set up form submission tracking so you can later separate agent-assisted conversions.
- Experiment in a staging environment. Ask developers to annotate one low-risk form with the declarative attributes and test it with the origin trial.
- Review security. Confirm server-side validation and rate limiting, and require human confirmation for purchases or account changes.
- Monitor the spec. Revisit quarterly; API names and browser support are still changing.
If your business sells online, read our guide to agentic commerce as well, because shopping is one of the first areas where agents are expected to act for users.
Common mistakes
- Treating WebMCP as an SEO ranking factor. It is not. It does not change how Google crawls or ranks pages.
- Shipping production code against an experimental API without feature detection. Always check that the API exists and expect names to change.
- Writing vague tool descriptions. "Submit form" tells an agent nothing; "Request a free roof inspection quote by postcode" does.
- Exposing sensitive actions without confirmation. Agents can misunderstand instructions; payments and deletions need a human step.
- Ignoring the basics. An inaccessible, slow, broken form will fail for agents and humans alike, with or without WebMCP.
Related Guides
- Reddit and AI Search: Why Forum Mentions Shape AI Answers
- Video and AI Search: Getting YouTube Content Cited by AI
- How to Get Your Brand Mentioned by Claude
Frequently Asked Questions
Is WebMCP an official web standard?
Not yet. As of October 2026 it is a draft being developed in the W3C Web Machine Learning group, with an origin trial in Chrome. Other browsers have not all committed to it.
Is WebMCP the same as MCP?
No. MCP connects AI applications to tools and data, usually on a server. WebMCP brings a similar tool idea into the web page itself, so browser agents can call actions using the user's existing session.
Does WebMCP help my Google rankings?
No. WebMCP is about letting AI agents complete tasks on your site. It is not a ranking signal, and Google's guidance for AI search features still centres on crawlable, helpful content.
Which AI agents use WebMCP today?
Chrome's documentation says support for Gemini in Chrome is coming, but no date was given. Independent tracking in mid-2026 found no mainstream agent using WebMCP tools in production.
Should small businesses do anything about WebMCP now?
Focus on well-labelled, accessible forms and clear calls to action. That helps people and today's screen-reading agents, and makes adopting WebMCP straightforward later.
Conclusion
WebMCP is one of the clearest signs that the web is being redesigned for AI agents as well as people. It is early, the API is still shifting, and you should not rebuild your site around it. But the preparation it rewards, clear actions, labelled forms and trustworthy flows, pays off immediately in usability and conversions.
Want a practical review of how agent-ready your site is? Our web development and conversion optimisation teams can audit your forms and key journeys, and our work on AI agents in marketing keeps us close to these standards. Get a quote to start.
References
- Chrome for Developers: Join the WebMCP origin trial
- Chrome for Developers: WebMCP Declarative API
- Patrick Brosset: WebMCP updates, clarifications, and next steps
- Spronta: The State of WebMCP, July 2026
- PPC Land: Chrome 149 origin trial puts WebMCP in developers' hands
- Semrush: WebMCP, what it is and what to do now
