Flipping it: Instead of AI living inside your apps, your apps now live inside AI! Instead of sending users to yet another tab, your application now lives directly inside the AI conversation. interactive, real-time, and context aware. In this session, we skip the theory and start with a live demo. You will see an MCP App running inside Copilot chat window, handling a real workflow without a single context switch. From there, we peel back the layers: what an MCP App is actually made of, how the host layer and your server divide responsibilities, and what the security primitives look like in practice. Then we get seriou: Authentication and SSO is the part that bites most teams, and we cover it properly. We walk through OAuth 2.1 with PKCE as the spec requires it, show how to wire enterprise SSO through Entra ID, and discuss what token scoping and consent flows look like when the window is owned by Microsoft as is in Microsoft Teams. We also cover the pitfalls: over-permissioned tokens, session handling across conversations, and prompt injection risks that come with rendering UI inside a chat interface. You will leave with a working mental model, a clear decision framework for what you must build versus what the platform handles, and enough practical knowledge to start building MCP Apps that are actually ready for production.

AI agents are only as trustworthy as the guardrails you build into them. In this hands-on session, we walk through the design and implementation of a custom Cowork skill that separates reading from acting, surfacing findings for human review before touching a single calendar, mailbox, or document. Using a real-world award-winning example from Blis Digital's proposal-to-project handoff, you will learn how to structure a skill in two distinct phases, write instructions that prevent the model from inventing data it cannot find, and handle edge cases like unknown external contacts gracefully. We will look at where the behavior lives and why a well-placed "are you sure?" does more for adoption than any clever automation. You will leave with a set of mental models for skill design, concrete instruction patterns you can apply immediately, and the confidence to pick that one recurring task you have been putting off automating.