MUP — Model UI Protocol

繁體中文

Bring interactive UI into LLM chat — so anyone can experience agentic AI, not just developers.

Current version: 0.2.11 | Protocol: mup/2026-03-17

Demos

Slides — AI builds a presentation live

Claude creates a full slide deck with charts, tables, and themes — all through function calls.

PDF to Presentation — AI reads a paper and presents it

Claude reads the AlexNet paper from a PDF, takes Markdown notes, captures figures with human-AI collaboration, builds a 19-slide presentation with charts and tables, then presents it slide by slide in the new in-panel reading mode — all narrated live.

Sound Pad — AI composes a track from scratch

16 browser-synthesized instruments, zero samples. Claude composes an electro swing track layer by layer, with event-driven narration synced to each section transition.

What is MUP?

A MUP is an interactive UI component that lives inside an LLM chat interface.

It bundles a visual interface with callable functions. The user operates it by clicking buttons; the LLM operates it through function calls. Both sides see each other's actions in real time.

The simplest MUP is a single .html file — no build step, no framework, no SDK.

Why?

Agentic AI is powerful, but today it's trapped behind text commands and developer tools. Most people never get to experience it.

MUP changes this. It puts clickable, visual UI right inside the chat — so anyone can use agentic capabilities without writing a single prompt.

Key Ideas

• Shared functions. A function can be called by the LLM (as a tool) or triggered by the user (via UI). Both sides operate on the same state through the same code.
• LLM as orchestrator. MUPs don't talk to each other. The LLM reads outputs and decides what to do next.
• Just HTML. Write a manifest, register your functions, done. Ship a single file.

Available MUPs

• Chat — Built-in, always available
• Slides — Presentation editor with charts, tables, themes, reading mode, and export (demo)
• PDF Reader — PDF viewer with page text extraction and region selection/capture
• Markdown — Markdown workspace with annotations and document management
• Sound Pad — 16-pad synthesizer with sequencer, per-track volume, and event system (demo)
• Chord Pad — Music theory keyboard with 5 synth voices, chord progressions, and melody playback
• Diagram — Auto-layouted node-and-edge diagrams (flowcharts, mind maps, architectures) with multiple layout algorithms
• Star Map — Location-based interactive star map with constellations, ecliptic, and Milky Way
• Voice — Speech synthesis and recognition (Jarvis mode)
• Notes — Sticky-note workspace with groups, styles, and export
• Pixel Art — Pixel canvas with shapes, ASCII patterns, fills, and PNG export
• Timer — Countdown timer / stopwatch with event-driven completion
• Progress — Task progress tracking
• More examples in archive/examples/ (games, productivity, etc.)

... View full README on GitHub