MCP Apps: Bringing Interactive UIs to AI Conversations (blog.fka.dev)

The Model Context Protocol (MCP) community has formalized on-demand UI generation with MCP Apps — an official extension (draft SEP-1865) and SDK that let MCP servers deliver rich, interactive HTML interfaces into AI conversation hosts (Claude, ChatGPT, Cursor, etc.). Instead of text-only outputs, servers can register UI Resources (ui:// URIs) and return content with UI metadata so hosts render mini web apps inline. This turns passive responses into interactive widgets with buttons, forms, charts and real‑time updates, improving usability for tasks like dashboards, data collection, and visual exploration. Technically, MCP Apps introduces three cooperating concepts: UI Resources (HTML templates, special MIME type like text/html+mcp), Tool–UI linkage (tools include a "ui/resourceUri" meta field to trigger rendering), and bidirectional communication via MCP’s JSON‑RPC over postMessage (e.g., tools/call). The repo includes types, examples and a TypeScript Node demo showing how to register a ui://counter/widget, expose tools (show_counter, get_counter, update_counter) with Zod schemas, and implement the client-side JS that posts JSON‑RPC requests to the host. By standardizing URIs, MIME handling and RPC messaging, MCP Apps makes interactive assistant UIs reproducible and portable across hosts, enabling richer, safer integrations and quicker developer iteration when building action‑oriented AI experiences.