What it does

MCP Apps is an extension to the Model Context Protocol that turns plain text tool responses into sandboxed HTML interfaces. A tool declares a ui:// resource; when the LLM calls it, the host fetches the HTML and drops it into a sandboxed iframe in the chat thread. The UI and host talk back and forth via postMessage, so the widget can both receive data and trigger other tools.

The interesting bit

The protocol is client-agnostic—same app runs in Claude, ChatGPT, VS Code, Goose, Postman, or MCPJam—but the repo itself ships no supported host implementation beyond a bare-bones example. The ecosystem bet is that chat clients will adopt the spec rather than each building proprietary plugin systems. There’s also a full “Agent Skills” scaffolding layer: install four skills and ask your coding agent to migrate an OpenAI app or wrap an existing web app.

Key highlights

• SDK split into four packages: core Views, React hooks, host app-bridge, and server registration helpers
• Examples cover genuinely varied terrain: CesiumJS globes, ShaderToy GLSL, sheet music rendering, cohort heatmaps, live transcription
• Includes migration path from OpenAI’s Apps SDK
• Apache 2.0, TypeScript, npm-published

Caveats

• Host support varies; the spec is an extension, not core MCP, and clients page is the source of truth for who actually implements it
• No production host implementation lives in this repo—hosts must build or adopt their own renderer (MCP-UI is one community option)

Verdict

Worth a look if you build MCP servers and are tired of explaining JSON output to users. Skip it if you just consume tools and your chat client of choice isn’t on the supported list yet.