Your website can't talk to AI agents yet. WebMCP changes that. Here's a plain-English guide to Google's new standard, and the Declarative vs Imperative APIs.
Key Takeaways
- WebMCP (Web Model Context Protocol) is a proposed Google web standard that lets AI agents interact with your website's features through explicit, machine-readable instructions instead of guessing from visual layout.
- The Declarative API lets you annotate existing HTML form elements with no new JavaScript — the right starting point for standard forms like search, contact, login, and checkout.
- The Imperative API uses JavaScript to programmatically register tools and handle invocations — built for multi-step wizards, dynamic filtering, and JS-heavy interactions.
- WebMCP runs in the browser and connects agents to client-side page tools; MCP runs on the server and connects models to backend data sources — a complete agentic architecture typically uses both.
- WebMCP itself is still experimental (Chrome 150+, origin trial), but its underlying principles — semantic HTML, structured data, llms.txt, and a clean accessibility tree — are worth implementing now.
If the agentic web has a technical foundation, WebMCP is it.
WebMCP (Web Model Context Protocol) is a proposed web standard from Google that lets AI agents interact with your website’s features in a structured, reliable way. Instead of agents guessing how to fill your forms or click your buttons by visually parsing the screen, WebMCP gives them explicit, machine-readable instructions.
HTML tells browsers how to render your page for humans. WebMCP tells AI agents how to use your page for tasks.
Why WebMCP Exists
Today, AI agents that browse the web face a fundamental problem: websites are designed for human eyes, not machine understanding. An agent looking at a flight booking form sees a wall of HTML — it has to guess which input is the departure city, which is the date, and which button submits the search.
WebMCP solves this by letting you explicitly declare: “this form is a flight search tool, this input takes a city name, this input takes a date, and this button runs the search.” The result is agents that are dramatically more reliable and faster, with far fewer misclicks and failed form submissions.
Two APIs, Two Approaches
The Declarative API — Annotate Your HTML
The simplest path. You add attributes to existing HTML form elements that describe what they do. No new JavaScript required.
When to use it: standard HTML forms — search, contact, login, checkout, filters. If your interaction starts with a <form> element, the Declarative API is your starting point.
The Imperative API — Register Tools with JavaScript
More powerful and flexible. You use JavaScript to programmatically register tools, define parameters, and handle invocations.
When to use it: complex interactions — multi-step wizards, dynamic filtering, real-time search, map interactions, or any functionality built primarily in JavaScript.
Declarative vs. Imperative: A Decision Framework
- Use Declarative if your interaction is form-based, your HTML is semantic, and the form does one clear thing
- Use Imperative if the interaction is JavaScript-heavy, involves multiple steps, or doesn’t use a traditional form
- Use both for complex pages — Declarative for forms, Imperative for everything else
WebMCP vs. MCP — What’s the Difference?
- MCP (Model Context Protocol) connects AI models to server-side data sources. It runs on the server.
- WebMCP connects AI agents to client-side web page tools. It runs in the browser.
MCP is for backend integrations, WebMCP is for frontend interactions. A complete agentic architecture might use MCP on the server and WebMCP on the client.
Should You Implement WebMCP Now?
Honest answer: not urgently, but prepare for it. WebMCP is experimental, requiring Chrome 150+ and an origin trial — which is also why it currently shows as “Not Applicable” in Lighthouse’s new Agentic Browsing audit for most sites. However, the underlying principles you should implement today are the same ones that audit already measures:
- Semantic HTML — proper form elements, labels, and ARIA attributes
- Structured data — Schema.org markup for content structure
- llms.txt — a machine-readable site summary at your domain root
- A clean accessibility tree — proper names and labels on every interactive element
The Bigger Picture
WebMCP represents a fundamental shift. For 30 years, we’ve built websites for human eyes. Now we need to build them for both humans and AI agents. This follows the same trajectory as mobile: first awareness, then best practices, then requirements. For AI agents, WebMCP is the responsive-design equivalent — and it’s arriving alongside a broader wave of agent-facing infrastructure, from Bing’s new AI visibility reports to Lighthouse’s agentic audits.
Frequently Asked Questions
What is WebMCP?
WebMCP (Web Model Context Protocol) is a proposed web standard from Google that lets AI agents interact with a website's features in a structured, machine-readable way, instead of visually guessing how forms and buttons work.
What's the difference between the Declarative and Imperative WebMCP APIs?
The Declarative API lets you annotate existing HTML form elements with attributes describing what they do, with no new JavaScript required — best for standard forms. The Imperative API uses JavaScript to programmatically register tools, define parameters, and handle invocations — best for multi-step, dynamic, or JS-heavy interactions.
How is WebMCP different from MCP?
MCP (Model Context Protocol) connects AI models to server-side data sources and runs on the server. WebMCP connects AI agents to client-side web page tools and runs in the browser. They're complementary — MCP for backend integrations, WebMCP for frontend interactions.
Do I need to implement WebMCP right now?
Not urgently. WebMCP is still experimental, requiring Chrome 150+ and an origin trial registration. But you should start implementing the underlying principles now: semantic HTML, Schema.org structured data, an llms.txt file, and a clean accessibility tree.
How does WebMCP relate to Google's Agentic Browsing score in Lighthouse?
Lighthouse's Agentic Browsing audit currently marks WebMCP-related checks as "Not Applicable" for most sites since the standard is experimental. But the three audits it does score today — llms.txt, accessibility tree quality, and layout stability — are exactly the foundational work that makes a future WebMCP implementation easier.
Written by mootkur9
17 years in digital marketing. Founder of Sadhyam. Writing about AI search, AEO/GEO, and the agentic web. More about me →
