Jeremiah Lowin explains how MCP apps let an agent bring an interactive interface into a conversation so a person can work directly with tools and server data. The design problem is narrower than building arbitrary consumer websites: Python developers often need tables, forms and charts for sharing or collecting information inside organizations.
Jeremiah Lowin presents Prefab as a way to compose existing interface components with Python classes and nested context managers. The Python description becomes a declarative JSON representation rendered by a React application. Reactive variables bind component values and client-side interactions, allowing Python authors to describe behavior without writing the corresponding JavaScript themselves.
Jeremiah Lowin illustrates progressively richer integrations with FastMCP: returning an interactive table from a tool, combining a table and chart in a grid, and defining an application with backend methods. A file-upload example shows why direct interaction matters: the user can send a file to the server through the app instead of having the agent reproduce its contents in a tool call. These capabilities require a client that supports MCP apps.
Jeremiah Lowin then describes generative interfaces based on the same structured representationFront-end code generation uses AI to create or modify the user-interface code of websites and applications from instructions or examples.. The recorded example streams agent-generated JSON into a renderer; a subsequently described approach streams Python, executes it in a sandboxAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access. and converts it to JSON. Jeremiah Lowin reports that the Python representation is about 70% smaller in this work, presenting potential token and latency benefits rather than an independently verified general benchmark.
Watch on YouTube
