The agent-ready web: Simplify user actions with WebMCP — Tara Agyemang, Google

What changed: Google’s Web MCP (Model Context Protocol) proposes a way for websites to expose first-class, schema-defined “tools” to in-browser agents, instead of forcing agents to parse DOMs and click pixels. It offers a declarative mode (HTML attributes on forms auto-generate JSON schemas) and an imperative mode (register custom tools with schemas and execute handlers). It’s in early preview, Chrome 146+ only, behind flags.

Why it matters: Agents get reliable, low-token access to site actions. For teams with complex, multi-step flows (booking, filtering, long forms), this can convert brittle automation into explicit, testable APIs embedded in the front-end. The shift is organizational: you stop hoping agents guess your UI and start owning a thin, tool-like surface for key flows.

Tradeoffs: You’re taking on API design inside your client. That means schema versioning, descriptive tool docs, and UI-state sync. Governance becomes mandatory: permission prompts, rate limiting, idempotency, audit logs, and payment interlocks (don’t let agents silently buy things). Abuse risk rises if tools make high-impact actions cheap to automate.

Adoption constraints: It only works with the browser open and currently targets Chrome + Gemini workflows. Cross-browser support and standardization are unclear. Claims of big reliability gains are demo-level, not production evidence. Security guidance, auth patterns, and bot-mitigation strategies weren’t detailed.

What most will miss: You’ll get 60–80% of the benefit by fixing fundamentals first—semantic HTML, accessibility, performance. Web MCP should sit atop that. Treat tool definitions like public APIs: stable names, error semantics, telemetry, and CI evals. Start with a single, high-friction flow; pilot, measure, and iterate.