Evolving AI chat with MCP Apps
summary
A proposed MCP Apps standard turns tools into sandboxed HTML/CSS/JS mini‑apps loaded in an iframe, using the @mcp/x‑apps RPC library to call server tools, push search results into the model context, and even request completions. Demo stacks include OpenRAG (Dockling + OpenSearch + Langflow), VS Code with GitHub Copilot and GPT‑5 mini, plus Goose and ChatGPT hosts; the search UI injects selected chunks back to the agent and works end‑to‑end though host support is still flaky. web‑powered interfaces enable richer shopping, booking and data exploration inside the chat.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 2:38AI and agent tools landscape
- 6:29Text-based AI limitations
- 9:17MCP vs skills for agents
- 11:03MCP server capabilities
- 13:53MCP apps for UIs
- 18:56Hello world MCP demo
- 23:53Advanced MCP with OpenRag
- 31:27MCP apps platform support
- 33:56MCP apps resources & future
- 37:49Q&A
related talks
22mMCP UI: Extending the frontier
26mBuild your First MCP App
29mMCP Apps: Primitives, discovery, and the Future of Software
23mIntroducing WebMCP: Agents in the Browser
20mLetting AI Interface with your App with MCP
32mProgramming language evolution in the AI era
15mWhen chatbots grow buttons: Building MCP apps
16mBuilding Interactive UIs in VS Code with MCP Apps