Pietro Zullo introduces MCP Apps as a way for an agent tool to return a useful interface rather than only text or JSON. The examples use UI resources and sandboxed widgets to let people interact with data, diagrams and controls inside an agent conversation.
Pietro Zullo explains how widgets and their host exchange context, stream arguments and make tool calls. The talk separates model-visible information from data intended for the interface, while emphasizing that application design still needs explicit decisions about context and behavior.
Pietro Zullo also discusses discovery, authentication, tool annotations and tests that help an app work in different clients. Client capabilities and interaction patterns vary, so a text fallback and capability checks remain important. The account describes the talk's architecture and demonstrations without presenting its release-period client list or distribution services as current guarantees.
Watch on YouTube
