Ruben Casas discusses the evolution of generative user interfaces for modern computing platform apps, highlighting the shift from static and declarative UIs to dynamic, AI-generated front-end code that enables highly adaptable and collaborative human-agent interactions. He envisions a future where MCP apps serve as secure, flexible environments for these generative interfaces, fostering innovative, personalized experiences beyond traditional component-based designs.
In this talk, Ruben Casas explores the evolution and future of generative user interfaces (UI) in the context of modern computing platforms (MCP) apps. He begins by reflecting on the early days of using ChatGPT to generate UI components through simple copy-pasting of code snippets, which he terms “poor man’s by coding.” Over the past few years, rapid advancements in AI models, particularly GPT-5.2 and Opus 4.5, have dramatically improved the quality and speed of UI generation, enabling models to produce high-fidelity, thoughtful, and accessible front-end code that can sometimes surpass human developers.
Despite these advancements, Casas questions why most UIs remain largely static and why we have not yet reached the “Jarvis moment” of dynamic, context-aware interfaces with floating windows and seamless interactions. He introduces the concept of the “new computer,” where interacting with AI models is akin to using a terminal without a mature graphical user interface. Two competing paradigms are emerging: embedding chat interfaces everywhere versus consolidating interactions within super apps like ChatGPT or Claude that render third-party UIs inside a unified environment. Both approaches are part of the ongoing search for the ideal interface for this new computing paradigm.
Casas then categorizes current UI generation methods into three types: static, declarative, and generative. Static UI involves predefined components with fixed props, similar to traditional web development. Declarative UI adds flexibility by using descriptors like JSON or YAML to dynamically generate UI layouts from static components, offering a balance between personalization and consistency. He cites examples like Netflix’s personalized UI and Vercel’s JSON Render as implementations of this approach. Declarative UI is seen as the current sweet spot for flexibility, speed, and cost-efficiency.
Looking forward, Casas envisions generative components as the next frontier, where AI models generate complete front-end code (HTML, CSS, JavaScript) on demand at runtime. While this approach offers maximum creativity and adaptability, it raises concerns about security and trust, necessitating sandboxing and containment strategies. MCP apps provide an ideal delivery mechanism for such generative UI due to their built-in authentication, tool calling, and sandboxing features. He highlights Anthropic’s use of MCP apps for first-party UI as a strategic example, suggesting that this protocol could become a standard for delivering dynamic, AI-generated interfaces.
Finally, Casas emphasizes that the future of generative UI will likely transcend static components toward collaborative human-agent experiences. He points to the Excalidraw MCP app as a pioneering example where humans and AI agents co-create and modify shared visual artifacts in real time. This collaborative model represents a new paradigm for interaction, moving beyond simple orchestration to a deeply personalized and interactive experience. While the ultimate form of UI in this new computing era remains uncertain, Casas encourages embracing creativity and shaping the future of interfaces together.