Beyond Components: Designing Generative UI for MCP Apps — Ruben Casas, Postman

TL;DR

• 1

  UI generation crossed a threshold with GPT-5.2 and Opus 4.5: Casas says the 2025 inflection point was not just better long-horizon reasoning but high-fidelity UI generation, including his own one-prompt blog rewrite that added a polished search box, blur animation, and accessibility without being asked.

• 2

  Chat everywhere is a temporary patch, not the final interface: He frames today's AI UX as the "terminal" era of a new computer, echoing Andrej Karpathy's idea that we have direct access to the system but have not yet invented the GUI.

• 3

  Most agent UIs still use static components dressed up with AI: In the common pattern, the agent just calls tools and passes props into prebuilt components, as seen in AGUI and Goose Auto Visualizer, which means the model is orchestrating data rather than inventing interface.

• 4

  Declarative UI is the practical sweet spot right now: Systems like Vercel's JSON Render let models generate JSON or YAML descriptors that map into a design system, giving more personalization while keeping consistency, predictability, speed, and lower token use.

• 5

  Generative components are the next step, but they need containment: Casas's Postman weather-agent experiment had a model generate HTML, CSS, JavaScript, and even a joke in one shot, but he argues this only works safely with boundaries and sandboxing, which is why MCP apps matter.

• 6

  The most interesting future is shared human-agent spaces, not floating Jarvis widgets: He points to the Excalidraw MCP app as a glimpse of where things go next, with a persistent canvas both human and agent can edit together instead of UI as one-way output.