What is WebMCP?

A browser-native protocol that lets websites declare structured tools for AI agents. It changes the relationship between your site and the agents visiting it, from observation to interaction.

April 2026

The short version

Today, AI agents interact with websites by reading HTML, parsing the DOM, or taking screenshots and using vision models to figure out what's on the page. They guess which buttons do what. They generate click actions and hope the page responds as expected. The full technical picture is covered in how AI systems actually browse the web.

WebMCP replaces guessing with declaration. A website states: here are the actions available on this page, here are the parameters each action requires, here is how to call them. An agent visiting the site discovers these declarations and calls them directly. No screenshots, no DOM parsing, no guessing. The interaction is structured, deterministic, and fast. Semrush described it as turning your website into an API that AI agents can use, without you having to build or maintain a separate API (Semrush, Mar 2026).

Where the idea comes from

In 2024, Anthropic released the Model Context Protocol (MCP), an open standard for connecting AI models to external tools and data sources. Before MCP, every integration between a model and an external tool was custom. MCP made it standard: tools declare themselves, the model reads the declarations, understands the parameters, and calls them directly. MCP was adopted quickly and is now supported across Claude, ChatGPT, Gemini, and most major AI development frameworks.

MCP operates at the application layer: it connects a model to backend systems. WebMCP applies the same concept at the browser layer: it connects an agent to a website's capabilities, mediated by the browser. The two protocols are complementary. A product might use MCP for headless backend operations and WebMCP for its customer-facing site.

How it works

The spec defines two approaches.

The first is a declarative approach for standard HTML forms. A website adds a few attributes to existing forms (a toolname, a tooldescription, an optional toolautosubmit flag), and the browser makes those forms discoverable by agents. A reservation form, a product search, a support ticket submission can all be made agent-accessible with minimal changes to existing HTML.

<form toolname="book_table"
      tooldescription="Book a restaurant table for a given date, time, and party size"
      toolautosubmit>
  <!-- The rest of your form stays the same -->
  <!-- Your form is now a callable WebMCP tool! -->
</form>

The second is a programmatic approach for complex, dynamic interactions: multi-step workflows, conditional logic, tools whose parameters depend on page state. A checkout tool that only appears when items are in the cart. A pricing tool that adjusts based on selections the user has already made.

navigator.modelContext.registerTool({
  name: "getCustomPrice",
  description: "Get a custom price offer for a product",
  inputSchema: {
    type: "object",
    properties: {
      "productId": {
        type: "string",
        description: "Product ID for custom offer"
      }
    },
    required: ["productId"]
  },
  execute: yourJavaScriptFunctionHere
});

// Your tool is now callable via WebMCP!

In both cases, the agent discovers what tools exist on the page, understands the schema, and calls them directly. One structured tool call replaces what used to require a long chain of browser actions.

Two directions of value

WebMCP matters for marketers because it unlocks value from two directions at once.

Discovery. Agents that can interact with your site through declared tools have reason to prefer your site over a competitor that requires screenshot-based navigation. Declared tools are faster, cheaper, and more reliable. If you are an agent choosing between two vendors, one whose site exposes structured tools and one that requires pixel-level guessing, the choice is obvious. The sites that are easiest for agents to work with will be the ones agents route users to.

Action. The piece most people are missing. WebMCP does not just help you get found. It lets agents act on your site. An AI agent researching vendors on behalf of a buyer can, through WebMCP, fill out a lead form, request a quote, check pricing, or start a trial natively. That is not a page view. That is a captured lead.

And an agent submitting structured data through a declared tool is a first-party intent signal. It enters your pipeline the same way any other qualified lead does. If you want to see what that looks like in practice, talk to us.

Where it stands today

WebMCP originated at Microsoft. Patrick Brosset on the Edge team published the initial proposal in August 2025. Google joined shortly after. By early 2026, the work had moved into the W3C, and the spec was published as a W3C Community Group Draft Report on February 10, 2026.

Chrome 146 has an experimental flag for testing. Google has published a live demo and a Chrome extension for inspecting registered tools (Chrome for Developers, Feb 2026). The spec editors are from Google and Microsoft. All four major browser vendors have representatives in the W3C community group. Microsoft Edge has expressed intent to support WebMCP but has not shipped anything. Firefox has begun syncing WebMCP Web Platform Tests, indicating active implementation work. Safari has no implementation.

The spec is a Community Group Draft Report, not a W3C Standard. The standard continues to actively evolve on a weekly basis through discussions on the webmachinelearning GitHub, with the most up-to-date information on the protocol often coming from the Chromium commit log itself. Chrome 146 is the only shipping implementation. We have an internal agent implementation running on a stagehand fork across a number of demo sites. Beyond Chrome's experimental flag and our work, we are not aware of other implementations in use.

Google's incentives

In March 2026, Google announced a dedicated user-agent string, "Google-Agent," specifically for when AI agents browse sites using Google infrastructure (Google developer docs). Project Mariner is a Chrome extension powered by Gemini. AI Mode keeps agent-style interaction inside Google Search. A2A routes agent-to-agent communication through Google Cloud. UCP, co-developed with Shopify, Walmart, and Target, standardizes commercial transactions for agents. Google's Head of Search, Liz Reid, recently described a future where "a lot of agents are talking with each other."

All of these share a common pattern: they keep agents interacting with the web through infrastructure Google controls. WebMCP fits the same pattern. It is a browser API that other browsers could implement, but Chrome is currently the only one that has.

There is also a historical pattern worth noting. Every major web standard Chrome has shipped has eventually become a Google Search ranking signal. HTTPS, mobile-friendly design, Core Web Vitals, structured data. We explore whether WebMCP might follow the same path in will WebMCP become a ranking signal.

The broader protocol landscape

WebMCP is not the only protocol in play. MCP handles backend tool integration. A2A handles agent-to-agent communication. UCP handles commercial transactions. WebMCP is the most credible candidate for browser-native, website-declared agent interaction specifically.

Whether it wins adoption at this point seems likely but is no guarantee. What is less uncertain is that some form of structured agent-site interaction protocol is coming. The incentive alignment between websites, agents, and browser vendors is too strong for the problem to go unsolved.

Why this matters now

The agentic web is not speculative. Google is shipping infrastructure for it. The major AI platforms are shipping agent features. Industry telemetry shows agentic traffic up 6,900% year-over-year. The volume of agents interacting with websites is growing rapidly.

WebMCP is one of the clearest signals of where the relationship between websites and AI is headed: from passive content consumption to structured, two-way interaction. The sites that declare their capabilities will be preferred by agents. The sites that let agents take action will capture leads and revenue that invisible sites never see.

If you're interested in learning more, talk to us.