modelcontextprotocol/ext-apps

What it solves

Standard MCP tools typically return text or structured data, which is limiting for complex visualizations. MCP Apps allows developers to build interactive user interfaces—such as charts, forms, dashboards, and design canvases—that render directly inside compliant AI chat clients like Claude and ChatGPT.

How it works

It extends the Model Context Protocol (MCP) by allowing tools to declare UI resources using a ui:// URI. When an LLM calls a tool, the host client fetches the associated HTML interface and displays it in a sandboxed iframe. This enables bidirectional communication where the host passes data to the UI via notifications and the UI can trigger other tools through the host.

Who it’s for

• App Developers: Those building interactive Views using the provided SDK and React hooks.
• Host Developers: Those embedding these Views into chat clients.
• MCP Server Authors: Those registering tools with UI metadata to enhance the user experience.

Highlights

• Broad Client Support: Compatible with Claude, ChatGPT, VS Code, Goose, and Postman.
• Agentic Workflow: Includes "Agent Skills" that allow AI coding agents to scaffold, migrate, or add UI to MCP servers automatically.
• Framework Agnostic: Provides starter templates for React, Vue, Svelte, Preact, Solid, and Vanilla JS.
• Extensive Examples: Ships with a wide range of demo apps including 3D globe viewers, real-time OS metrics, and PDF viewers.