MCP tools graduate from text walls to interactive UIs

What it does

MCP Apps extends the Model Context Protocol so that tool definitions can carry a ui:// resource—an HTML interface instead of a plain text response. When a host like Claude or ChatGPT calls the tool, it fetches that resource and renders it in a sandboxed iframe inside the chat thread. The repo contains the spec, a TypeScript SDK for building and embedding these Views, and a broad set of example servers.

The interesting bit

The protocol effectively treats chat clients as a lightweight windowing system: the UI runs isolated in an iframe but can call back into the host to trigger other tools, while the host pushes updates via notifications. It is a pragmatic way to avoid rebuilding chat infrastructure just to show a spreadsheet or a 3D globe.

Key highlights

• Ships a four-package SDK covering app Views, React hooks, host embedding (app-bridge), and server registration.
• Bundles four “Agent Skills” to scaffold new apps, migrate OpenAI Apps, add UI to existing servers, or convert web apps.
• Claims client support across Claude, ChatGPT, VS Code, Goose, Postman, and MCPJam, though the README warns that host support varies.
• Includes example servers ranging from Excalidraw and CesiumJS maps to budget allocators and cohort heatmaps.
• No supported host implementation lives in this repo beyond a basic-host example; production clients appear to rely on the external MCP-UI framework.

Caveats

• Host support is explicitly uneven; the spec is an extension, and not every MCP client renders ui:// resources.
• The repo does not ship a production-ready host implementation, so client authors must either adopt MCP-UI or build their own iframe orchestration.

Verdict

Worth a look if you build MCP servers and want to stop explaining data with monospace text dumps. Skip it if you are looking for a drop-in chat client—this is the server and spec side of the equation.