Google's WebMCP — The Web as a Structured Database for AI Agents

Date: 2026-02-16 | Source: The Decoder

TL;DR

Google proposes WebMCP — bringing the Model Context Protocol (MCP) to websites so AI agents can interact with them via structured tools (book flights, create tickets, search products) instead of parsing raw HTML.

What Is WebMCP?

• Google's adaptation of MCP specifically for the web/browser
• Websites define structured tools that AI agents can call directly
• Two APIs:
  • Declarative API — simple actions via HTML forms
  • Imperative API — complex processes via JavaScript
• Makes websites "agent-ready" — agents interact with structured data, not raw page code

Use Cases

• E-commerce: Agents search, filter, configure options, navigate checkout
• Travel: Search flights, filter results, handle bookings with precision
• Customer support: Auto-fill detailed support tickets

Timeline & Reality

Current status: Early Preview Program (developer access only via Google's EPP)

When will it be real?

• 🔵 Now: Developer preview — experimental, limited access
• 🟡 2026 H2 (estimated): Chrome integration likely, Gemini agents could use it
• 🔴 Full adoption: 1-2+ years — requires websites to implement WebMCP APIs

Key blockers:

1. Security: Google explicitly says prompt injection defense is the agent's responsibility, not the API's. OpenAI admits prompt injection "may never be fully solved."
2. Adoption: Websites must actively implement WebMCP — same chicken-and-egg problem as any web standard
3. Business model conflict: If agents handle everything, users don't visit websites → lost ad revenue, customer contact

Competition

• Microsoft NLWeb (mid-2025): Open-source, natural language interface for websites, each instance acts as MCP server
• Cloudflare Markdown for Agents: Simpler approach — just serve markdown to agents

Impact Assessment

• The "agentic web" is still more vision than reality
• Current practical answer: heavily constrained agents with human oversight
• For Kinedu: worth monitoring but don't invest in implementation yet — wait for Chrome integration and broader adoption signals
• The real question: do you want Kinedu to be "agent-ready" (discoverable by AI) or resist agent access (protect direct traffic)?