What if your website could tell an AI agent exactly what it can do - subscribe to a newsletter, search an article, fill out a form - without the agent having to scrape HTML or call a backend API?
That's the idea behind WebMCP, an emerging approach being discussed in a W3C community group. It builds on Anthropic's Model Context Protocol (MCP), the open standard for connecting AI tools to data sources, and brings it into the browser itself.
Developer Richard MacManus just documented his experience adding WebMCP support to his WordPress site. He registered two tools using a JavaScript interface called navigator.modelContext.registerTool(): one that subscribes an email address to his newsletter, and another that searches the current page for a relevant paragraph and highlights it visually. Both run directly in the browser with full DOM access - no server-side integration required.
Two Ways to Try It Today
There are currently two implementation paths:
- MCP-B (the "B" stands for browser) - a browser extension that acts as a bridge between AI agents and web page tools. Works in Chrome, Edge, and Firefox right now.
- Chrome Canary's native WebMCP - Google shipped an experimental implementation last month, building MCP support directly into the browser without needing an extension.
The native approach is the more interesting one long-term. Instead of relying on an extension as middleman, the browser itself mediates between AI agents and website-registered tools. Think of it as similar to how browsers standardized geolocation, notifications, and local storage - WebMCP could become the standard layer for AI interaction.
What This Means for Site Owners
Today, AI agents interact with websites through clumsy workarounds: scraping page content, parsing HTML structures that can change without warning, or requiring custom API integrations. WebMCP flips this by letting site owners explicitly declare what their site can do for an agent.
The practical applications go beyond simple demos. E-commerce sites could expose product search and cart functionality. Documentation sites could let agents query specific sections. SaaS tools could surface their most common actions without building separate API endpoints.
This is still early. WebMCP is not a formal web standard yet, and the Chrome Canary implementation is explicitly experimental. But the W3C community group's involvement signals this could follow the path from experiment to specification to widespread adoption. For developers already building MCP integrations on the server side, the browser-native version is a natural extension worth tracking.