Chrome 146 Introduces WebMCP: Websites Become Callable Functions for AI Agents
Key Takeaways
- ▸Chrome 146 Canary includes WebMCP, a new API allowing websites to register callable tools for AI agents with structured inputs/outputs
- ▸Developed jointly by Google and Microsoft with W3C draft published February 10, 2026; stable Chrome release expected March 2026
- ▸Real-world implementation on snout.pink demonstrates the concept works today, with 6 tools built in 456 lines of code in one afternoon
Summary
Google has quietly introduced WebMCP (Web Model Context Protocol) in Chrome 146 Canary, a new browser API that fundamentally changes how AI agents interact with websites. Instead of screen-scraping and DOM parsing, websites can now register structured tools that agents can call directly as functions. The API allows sites to expose their capabilities through typed parameters and return structured JSON responses, similar to traditional API calls but native to the browser. Developed jointly by Google and Microsoft, the W3C draft was published on February 10, 2026, with the stable Chrome release expected in March 2026.
Developer gurkenkoenig demonstrated the technology by implementing WebMCP on their subscription tracker snout.pink, creating six callable tools (list, add, update, delete, summary, export) in just 456 lines of code written in an afternoon. The implementation uses navigator.modelContext.registerTool() to expose website functions with formal schemas defining inputs and outputs. AI agents connecting via Chrome DevTools Protocol can discover and invoke these tools without rendering the user interface, enabling direct structured communication between agents and web services.
The implications are far-reaching: AI agents could handle tasks like hotel bookings through simple function calls (search_hotels → check_availability → book_room) rather than navigating complex UIs filled with dark patterns and advertisements. News sites could expose articles through tools like get_latest_articles and search_by_topic, allowing agents to filter and summarize content without users ever seeing homepages or clickbait. This represents a fundamental architectural shift where websites evolve from visual interfaces designed for human interaction into structured APIs for AI consumption, potentially rendering the traditional web browsing experience obsolete.
- AI agents can discover and call website functions directly via Chrome DevTools Protocol without rendering UI or screen-scraping
- Technology could fundamentally transform web interaction from visual interfaces for humans to structured APIs for AI agents
Editorial Opinion
WebMCP represents a fascinating inflection point where the web's presentation layer becomes optional infrastructure. While the efficiency gains for AI agents are undeniable, this shift raises profound questions about the web's open ecosystem: if websites become callable functions optimized for AI consumption, what happens to the visual web that humans can inspect, share, and understand? The technology also creates new centralization pressures—Google and Microsoft controlling the standard through which AI agents access web services could give them unprecedented leverage over both content providers and competing AI platforms. Most critically, as the author notes, this isn't speculative: it's shipping in weeks.
