In this talk, Pietro Zullo, co-founder of Manufact, Inc., introduces MCP apps, explaining their fundamentals, how they function, and their distribution mechanisms. MCP (Model Context Protocol) has been evolving since 2024, with MCP apps emerging around late 2025 as an extension that allows MCP servers to return interactive UI components rather than just JSON. This evolution enables richer user experiences by embedding sandboxed iframes within chat interfaces, facilitating bidirectional communication between the UI and the host application. Pietro emphasizes that MCP apps represent the future of software interaction, where models not only process data but also present dynamic, interactive interfaces.
Pietro outlines the core primitives of MCP apps, such as the ability to update the model’s state based on UI interactions, send follow-up messages from the UI to the chat model, and stream partial inputs to dynamically update UI components in real time. These features enable sophisticated interactions, like rendering SVG graphics or live video within the chat interface. Additionally, MCP apps support calling other tools from within the UI, enhancing modularity and interactivity. Privacy considerations are addressed by allowing MCP servers to separate data shown in the UI from data sent to the model, ensuring sensitive information can be displayed to users without exposing it to external models.
The talk also covers client support and development tools. Several popular clients, including ChatGPT, Claude, and Cursor, support MCP apps with varying degrees of functionality. Manufact provides open-source SDKs, such as MCPUs, which simplify building MCP servers and apps by abstracting protocol complexities. Developers can create UI widgets using React components compiled into HTML and CSS, facilitating rapid development. Pietro highlights that MCP servers can detect whether the client supports MCP apps and adjust their output accordingly, ensuring compatibility across different environments.
Distribution and discovery of MCP apps are critical topics addressed in the presentation. MCP app stores have been launched by major providers like ChatGPT, Claude, and Cursor, offering self-serve submission processes that vet and publish MCP servers and apps. Once published, apps become discoverable by millions of users and can be installed with a single click, eliminating the need to share complex configuration files. Notably, Cloud currently supports dynamic discovery, allowing it to automatically find and use the most appropriate MCP connectors based on user requests, a feature expected to be adopted soon by ChatGPT as well. This dynamic discovery promises significant organic user growth for MCP app developers.
In conclusion, Pietro stresses the transformative potential of MCP apps, likening AI apps to the new browsers and platforms like ChatGPT to new websites that deliver interactive experiences directly within chat interfaces. He shares personal experiences of integrating MCP apps into workflows, demonstrating how these apps streamline tasks by connecting various tools and data sources seamlessly. The talk encourages developers to embrace MCP apps as a way to reach users more effectively and build the next generation of software experiences embedded in AI-driven environments. Pietro offers support and resources for those interested in building and distributing MCP apps, underscoring the growing ecosystem and opportunities in this space.
