Daniel Rosehill

A New Protocol for the Agentic Web

WebMCP is a proposed web standard that enables websites to expose structured tools to AI agents operating within the browser. It is being incubated in the W3C Web Machine Learning Community Group, with editors from Google and Microsoft. Chrome opened an early preview in February 2026 and moved it into a public origin trial from Chrome 149. It is not a finished standard yet, but the implications are already significant.

The core idea: instead of AI agents having to visually interpret a webpage (reading pixels, guessing at button locations, parsing unstructured HTML), websites can declare exactly what actions are available, what inputs they expect, and what outputs they return. The agent doesn't need to "see" the page. It can interact with it directly through a structured interface.

How It Works

Chrome's early preview introduced two complementary ways for a site to expose tools:

Declarative API

Handles standard actions defined directly in HTML forms. If your site already has well-structured forms — search fields, filters, checkout flows — the Declarative API lets you expose those as agent-callable tools with minimal additional work.

Imperative API

Manages complex, dynamic interactions that require JavaScript execution. For SPAs, interactive configurators, or multi-step flows that go beyond simple form submissions, the Imperative API provides the programmatic interface agents need.

Together, these cover the full spectrum of web interactions — from simple search queries to complex multi-step transactions. The imperative side is the more settled of the two: the current community-group draft specifies it as document.modelContext.registerTool(), where each tool has a name, a description, a JSON Schema for its inputs, and a function that runs it.

Why This Matters: The Problem with Computer Vision

Today, most AI agents that interact with websites rely on computer vision — they take screenshots, identify UI elements visually, and simulate mouse clicks and keyboard input. This approach is fundamentally brittle:

  • Layout changes break everything. A redesign, an A/B test, even a minor CSS change can throw off element detection.
  • Dynamic content is unreliable. Loading spinners, lazy-loaded images, and JavaScript-rendered content all create timing issues that visual agents struggle with.
  • It's slow. Screenshot, process, identify element, simulate click, wait, screenshot again. Every interaction involves multiple round trips through a vision model.
  • It doesn't scale. Each site requires its own visual understanding. There's no standard to rely on.

WebMCP replaces all of this with direct, structured communication. The agent asks "what tools are available?" and the website responds with a machine-readable list of actions, parameters, and expected outputs. No screenshots. No pixel-hunting. No guessing.

The E-Commerce Opportunity

This is where the business case becomes compelling. E-commerce is one of the primary use cases Google highlights, and for good reason.

Consider what an agent-equipped shopper looks like in practice: "Find me a navy blazer under $200, size 42 regular, available for delivery by Friday." Today, an AI agent handling that request would need to navigate to a retailer's site, visually identify the search bar, type a query, parse the results page, apply filters by clicking dropdowns, and scrape product details from the DOM.

With WebMCP, the retailer exposes structured tools: search_inventory, filter_products, check_availability, add_to_cart. The agent calls these directly with typed parameters and gets structured data back. Faster, more reliable, and — critically — under the retailer's control.

Competitive Advantage for Early Adopters

Retailers who implement WebMCP early will be the ones AI agents can actually interact with reliably. When a customer's agent is comparison shopping, the sites with structured tool interfaces will get accurate results returned. The sites still relying on visual scraping will get skipped, misrepresented, or fail entirely.

This is a new kind of SEO — not for search engines, but for AI agents.

Beyond E-Commerce

Google's early material highlighted commerce, travel and support, but the applications extend further:

Customer Support

Agents can auto-populate technical details in support tickets, navigate knowledge bases, and escalate with full context — without scraping chat widgets.

Travel

Flight searches, hotel filtering, booking flows, and itinerary management through structured tools instead of brittle form automation.

Financial Services

Account lookups, transaction queries, and form-heavy applications like loan pre-qualification — all exposed as typed, validated tool interfaces.

Government & Public Services

Permit applications, benefits enrollment, and document submission — complex bureaucratic flows that agents could navigate on behalf of citizens.

Current Status

As of October 2026, WebMCP is in a Chrome origin trial, starting with Chrome 149. Sites can register for a trial token and switch the API on for real visitors, with no browser flags required. At Google I/O in May 2026, Google said Gemini in Chrome will "soon" support WebMCP. Brands including Expedia, Booking.com, Shopify, Etsy, Instacart and Target were shown as experimenting with it.

The specification itself is still a draft community-group report, not a W3C standard, and the API has already changed shape once — Chrome's early preview used navigator.modelContext, while the current draft uses document.modelContext. Anything built today should expect to be adjusted.

It is still early — but the direction is clear. Browsers are being given the infrastructure for websites to become first-class participants in the agentic web, not passive targets of screen-scraping bots.

What This Means for Builders

If you're building or maintaining a web application — especially one with transactional flows — WebMCP is worth watching closely. The sites that expose structured tool interfaces will be the ones AI agents interact with most effectively. The sites that don't will increasingly be left out of agent-mediated workflows.

This site practises what it describes: every page registers a small set of read-only WebMCP tools — a site overview, a page list, a page reader and contact options — so a browser agent can find its way around without scraping. Nothing that submits data is exposed.

The protocol also aligns naturally with the MCP servers I build for backend systems. A WebMCP-enabled frontend paired with an Admin MCP backend creates a complete agentic surface — public tools for customer-facing agents, admin tools for internal operations, all speaking the same structured protocol.