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

Ruben Casas, Staff Engineer at Postman, argues that while frontier models such as GPT-4.5 and Claude Opus 4.5 (released in late 2025) have reached the point where they produce higher-fidelity frontend code than most human engineers, the UIs surfaced through AI agents remain largely static and pre-built. He frames this gap as the central design challenge of the current moment and uses the talk to map a three-level spectrum for generative UI, then explains why MCP apps are the correct distribution layer for the most advanced tier.

Casas opens by recounting his own inflection-point moment: he prompted a model with a single instruction to rewrite his blog, and without being asked, it generated a search box with a blur animation and built-in accessibility. That unprompted quality led him to concede the model writes better frontend code than he does. He notes that the wider acceleration visible in the final months of 2025 — driven largely by GPT-4.5 and Opus 4.5 — made high-fidelity, long-horizon UI generation a practical reality.

The talk then surveys three levels of UI generation. Level one is static components: the agent acts as an orchestrator, makes a tool call via MCP or direct tool invocation, passes props and data to predefined React components, and the client renders them. Examples include the AG-UI protocol (an SDK that maps tool-call parameters to registered React components) and Goose's Auto Visualizer feature (which matches arbitrary data to a curated component library). Casas notes this model is functionally identical to the server-sends-data / client-renders-component pattern developers have used for two decades.

Level two is declarative UI: the LLM still targets a fixed component library but instead of passing raw props, it generates a JSON, YAML, or Python descriptor that a translation/rendering engine maps to components at runtime. Casas draws a parallel to Netflix's server-driven UI, which has personalized homepages against a fixed component palette since the early personalization era. He highlights Vercel's JSON Render (recently extended with YAML support) as a contemporary example. He argues declarative UI represents the best practical balance today — it preserves design-system consistency and predictability, reduces token consumption relative to full code generation, and enables far more dynamic, personalized layouts than static props alone.

Level three is fully generative components: the model writes HTML, CSS, and JavaScript on demand at runtime, bypassing any predefined component library entirely. Casas demonstrates this with a weather agent he built at Postman that, in a single tool call, fetches weather data, generates a joke, and produces a complete, novel HTML/CSS/JavaScript UI rendered in the browser — no components, no translation layer, purely LLM-authored code. The core challenge here is trust and containment: LLM-generated code must be sandboxed before being presented to users. This is precisely why Casas argues MCP apps are the ideal delivery mechanism for generative UI — they provide authentication, tool calling, message-passing between UI and agent, and, critically, sandboxed execution via a double-iframe default. He notes that Anthropic's own decision to use MCP apps for Claude's first-party visualizer feature (rather than building a proprietary rendering architecture) is a strong signal that the protocol is the right standard.

Casas closes with the TV-and-radio analogy: the first television shows were radio shows filmed with cameras because nobody could yet imagine what else to do with the new medium. He argues we are in the equivalent era for agent interfaces — chat everywhere and MCP super-apps are plausible but almost certainly not the final form. The more interesting future, he suggests, lies in collaborative human-agent artifacts exemplified by the Excalidraw MCP app, which creates a shared canvas where a human and an agent can modify the same diagram in real time. Casas frames this as the direction beyond components: not just generative output, but genuinely collaborative, personalized, co-created UI experiences.