Prefab — Generative UI Framework

Status: Initial research entry
Last updated: April 9, 2026
Maturity: Alpha (Prefect)

Why This Matters to This Research

Prefab is the first production-oriented framework that strongly validates the declarative, component-tree approach to UI that I've been exploring through YAML-driven platforms and MCPApps. It shifts the authoring layer to a clean Python DSL while producing a framework-agnostic JSON spec — exactly the kind of constrained, composable UI that agents can reliably generate and deploy.

Discovery

Announced April 8, 2026 by Jeremiah Lowin (Prefect CEO). Discovered April 9, 2026.

Tagline: "Composition, not construction."

Assemble using 100+ prebuilt, high-quality components rather than building new ones from scratch in Python.

Core Architecture

Python DSL (authoring)
    → JSON component spec (framework-agnostic)
    → Bundled React renderer (shadcn/ui)

Python is used only for authoring.
The output JSON spec contains no Python references.
This enables agents, browsers, and other tools to work with the same intermediate format.

Installation:

pip install prefab-ui
Python 3.10+, Apache 2.0 license.
Still in alpha — expect breaking changes.

Component Library

Prefab ships with 100+ production-quality components built on shadcn/ui, providing consistent design, accessibility, and polish out of the box.

Category	Components
Layout	Grid, Column, Row, GridItem
Cards & Containers	Card, CardHeader, CardTitle, CardDescription, CardContent, CardFooter
Data Display	DataTable, DataTableColumn, Badge, Progress, Ring, Metric, Sparkline
Charts	BarChart, RadarChart, Sparkline with ChartSeries
Inputs & Controls	Button, Checkbox, Radio, RadioGroup, Switch, Slider, Combobox, DatePicker
Feedback	Alert, AlertTitle, AlertDescription, Tooltip, HoverCard, Dialog, Loader
Control Flow	ForEach, If, Elif, Else, Condition, SetInterval
Effects	Carousel, Separator

Reactive State & Interactivity

Client-side reactivity is handled entirely through the Rx class — no custom JavaScript required.

Template expressions: {{ ctx_tick % 20 * 3 + 20 }} with support for math, pipes (currency, percent, truncate), and conditionals.
Actions: SetState, ShowToast, on_click, on_change, on_success.
Result: Full interactivity with zero manual JS.

FastMCP Integration

Prefab has native, first-class support for FastMCP with three practical modes:

Interactive tools — Add app=True to any tool and return Prefab components directly.
FastMCP Apps — Dedicated app class with clear separation: @app.ui() for UI and @app.tool() for backend logic.
Generative UI — LLMs write Prefab Python code, which executes inside a sandbox and renders dynamically.

Deployment

prefab export examples/app.py generates a self-contained static bundle.
No server required. Easy to embed in iframes or host on Vercel or any static host.
This provides a clean path from MCP tool → fully deployed interactive app.

Relation to Existing Research Threads

Prefab directly strengthens two parallel lines of exploration:

YAML-Driven UI Platform

Same core idea (compile to a component tree) but with a different authoring layer. Prefab brings a mature shadcn-based component library and polished React renderer to the pattern. The YAML platform still leads in runtime features (full Juris state interpolation, dual-protocol support) developed a year earlier.

Vercel Apps / MCPApps

prefab export delivers the exact static deployment story needed. The generative mode (LLM → sandboxed Prefab → browser) is the concrete realization of agent-driven UI generation.

Strengths & Current Limitations

Strengths:

Excellent out-of-the-box component quality and consistency (shadcn/ui baseline)
Clean separation between authoring layer (Python) and runtime (JSON spec)
Strong static export story and native FastMCP integration
Agent-friendly constrained format — LLMs can generate valid UIs without knowing React

Limitations (as of April 2026):

Alpha stage with expected breaking changes; pin versions for any real usage
Python-only authoring layer — no native YAML mode yet
Sandbox security model, performance with large component trees, and escape hatches to custom React components are still emerging
Limited public examples and documentation at launch

Open Questions & Future Research Directions

Could Prefab's JSON spec become a shared interchange format between Python DSL, YAML, and LLM-generated UIs?
How well does the sandbox scale for complex, multi-step agent-generated interfaces?
What would a hybrid YAML + Python authoring mode look like?
Long-term: Could Prefab evolve into the standard "React for agents"?