Most “chat with our app” demos ignore the one thing users actually care about: what’s on this screen, this form, this dataset right now. In this session, we’ll build a context-aware agent for a web app that behaves more like R2-D2 than a tab-open LLM—aware of the current page, able to take safe actions, and honest about what it can’t do. We’ll wire an agent to read structured context from the DOM and client-side state, then expose a small set of tools for it to call: validate inputs, suggest changes, trigger workflows.