Summary
Jeremiah Lowin, creator of FastMCP and founder of Prefect, presents a novel approach to building interactive UIs for Python developers working within the MCP (Model Context Protocol) ecosystem. The talk begins with an explanation of MCP apps — an extension introduced in early 2026 that allows AI agents to serve full HTML/CSS/JavaScript interfaces directly to users, bypassing the agent’s context window for richer, stateful interactions like booking tables or displaying schedules.
The core problem Lowin addresses is that FastMCP’s primarily Python-based user base needs frontends without entering the JavaScript ecosystem. The solution is Prefab: a Python DSL that compiles to a JSON intermediate representation, rendered by an underlying React application. The key architectural insight is that the JSON layer is serializable — meaning it can be generated by an agent, modified by an agent, sent to a human for editing, or produced programmatically and rendered live. Prefab ships with 130+ composable components, and its own documentation is rendered entirely within Prefab itself, with live-editable Python code examples.
Lowin frames Prefab explicitly for enterprise Python teams who need to share and collect structured information internally — not for building consumer-grade branded frontends. An upcoming MCP protocol update will allow agents themselves to interact with MCP app UIs, which Lowin suggests will enable use cases like agent-vs-human chess in a visual interface. The talk is essential viewing for Python engineers building agentic workflows who need to surface data or collect input without leaving their native stack.
📺 Source: AI Engineer · Published September 10, 2026
🏷️ Format: Hands On Build
