AI agents can browse the web but they don’t truly understand it.

Most current solutions rely on screen scraping or vision models that interpret UIs as images. The result is slow, fragile systems that break on dynamic interfaces. Custom backend integrations improve control but introduce setup complexity and eliminate real-time discoverability.

What if the web itself could expose structured capabilities directly to AI agents?
In this talk, we introduce WebMCP, a new approach where web applications define native, structured tool contracts inside the browser, enabling agents to interact with them reliably without scraping or custom integrations.

You will see how agents can discover tools on the fly, execute actions through JSON schemas, and stay in sync with application state, all mediated securely by the browser.

We will walk through real implementation patterns using Reactjs, explore how to design effective tool contracts, and cover authentication, debugging, and production considerations.

If you’re building modern web apps or working with AI agents, this session will give you a practical blueprint to move from fragile automation to scalable, AI-native web experiences.

What participants build: a working set of WebMCP tools registered on the starter app, plus one agent-driven task run end to end.

What they take home: a reusable pattern for turning any app action into an agent-callable tool, and a checklist for deciding which actions are safe to expose.

Prerequisites: basic React, JavaScript or TypeScript