model-compose

One YAML file. Any model. Any protocol. Any runtime.

Build chat APIs, RAG pipelines, autonomous agents, and MCP servers without writing orchestration code — then run the same definition anywhere, like docker-compose.

AI systems shouldn't be locked into a single provider, runtime, or cloud. model-compose is built on four principles:

• Composable — Models, agents, workflows, tools, memory, and protocols are interchangeable building blocks.
• Portable — Define your AI system once and run it anywhere without re-engineering.
• Model-Agnostic — Mix local models and cloud APIs in the same system.
• Stream-Native — Tokens, audio, video frames, and events flow through workflows as first-class values.

Quick Start

Install with uv:

uv pip install model-compose

Or with pip:

pip install model-compose

Create model-compose.yml:

controller:
  adapter:
    type: http-server
    port: 8080
  webui:
    port: 8081
workflow:
  job:
    component: chatgpt
    input:
      prompt: ${input.prompt}
    output: ${output as stream/text}
component:
  id: chatgpt
  type: http-client
  base_url: https://api.openai.com/v1
  action:
    path: /chat/completions
    method: POST
    headers:
      Authorization: Bearer ${env.OPENAI_API_KEY}
      Content-Type: application/json
    body:
      model: gpt-4o
      messages:
        - role: user
          content: ${input.prompt}
      stream: true
    stream_format: json
    output: ${response[].choices[0].delta.content}

Run it:

export OPENAI_API_KEY=your-key
model-compose up

That's it. You're serving GPT-4o at http://localhost:8080 with a web UI at http://localhost:8081. No application code. No framework boilerplate. Same file runs locally, in Docker, or in production.

Automatic Web UI

Every workflow gets a browser UI for free — inputs, outputs, streaming, and file uploads are inferred from the schema. Two lines of YAML, no frontend code.

controller:
  webui:
    driver: gradio
    port: 8081

!Automatic Web UI

Perfect for iteration, demos, and sharing a workflow with non-engineers before wiring it into production.

What You Can Build

Here's what a single YAML file can serve today — just a few examples.

🤖 Autonomous Agents

Build a ReAct agent that plans, uses tools, and completes multi-step tasks — declaratively.

component:
  id: research-agent
  type: agent
  tools: [search-web, fetch-page]
  max_iteration_count: 10
  action:
    model:
      component: chatgpt
    system_prompt: You are a web research assistant.
    user_prompt: ${input.question}

See simple agents like a code reviewer, a RAG assistant, and a web researcher in agents/.

🧠 Local Models

Run open-weight models on your own hardware with PyTorch, vLLM, or llama.cpp — same YAML surface as cloud APIs, no SDK code.

component:
  type: model
  task: chat-completion
  model: HuggingFaceTB/SmolLM3-3B
  action:
    messages:
      - role: system
        content: ${input.system_prompt}
      - role: user
        content: ${input.user_prompt}

Point model: at any Hugging Face repo, or a local GGUF file via llama.cpp. Browse local chat, embedding, TTS, VLM, image/audio/video tasks under model-tasks/.

🔍 RAG Pipelines

Compose embedding, vector search, and generation into a single workflow — no glue code.

workflow:
  jobs:
    - id: embed
      component: embedder
      input: { text: ${input.query} }
- id: retrieve
      component: knowledge
      action: search
      input: { vector: ${jobs.embed.output} }
- id: answer
      component: chatgpt
      input:
        context: ${jobs.retrieve.output}
        question: ${input.query}

Native drivers ship for Chroma, Milvus, Qdrant, FAISS, Neo4j, ArangoDB, and Redis.

🎬 Media Pipelines

Chain audio, video, and image stages across models and native codecs — swap any step without touching the rest.

workflow:
  jobs:
    - id: extract-audio
      component: audio-extractor
      input: { source: ${input.video as video} }
- id: transcribe
      component: whisper
      input: { audio: ${jobs.extract-audio.output} }
- id: translate
      component: translator
      input: { text: ${jobs.transcribe.output} }
- id: dub
      component: tts
      input: { text: ${jobs.translate.output} }
- id: lip-sync
      component: wav2lip
      input:
        video: ${input.video as video}
        audio: ${jobs.dub.output as audio/wav}
      output: ${output as video/mp4}

See full video dubbing, scene detection, voice cloning, and face mosaic pipelines in media-processing/.

⚡ Streaming Multi-Modal Workflows

Stream tokens, audio chunks, and video frames end-to-end — first-class across every stage.

workflow:
  job:
    component: chatgpt
    output: ${output as sse-text}
component:
  id: chatgpt
  type: http-client
  action:
    body: { stream: true, ... }
    stream_format: json
    output: ${response[].choices[0].delta.content}

Real-time TTS, video-to-frames, and live chat examples live under data-streaming/ and showcase/.

🌐 MCP Servers

Turn any workflow into an MCP server that Claude, ChatGPT, or Cursor can use — one line change.

controller:
  adapter:
    type: mcp-server   # ← was: http-server
    port: 8080

Full examples live in mcp-servers/, including a Slack bot MCP.

From Development to Production

The same YAML that runs on your laptop scales without a rewrite.

1. Develop locally

model-compose up

Runs on your machine with a Gradio web UI at :8081 — perfect for iteration.

2. Deploy as a container

Add a runtime: block. Same file, same behavior:

controller:
  runtime:
    type: docker
    image: my-ai-service:latest
    ports: [ "8080:8080" ]

3. Scale horizontally

Add a queue. Dispatchers accept jobs, subscribers process them across N machines:

controller:
  adapter: { type: http-server, port: 8080 }
  queue:
    driver: redis
    host: redis.internal
    name: my-queue

No shared filesystem. No code changes. Just add more subscribers to scale.

Why model-compose?

| | model-compose | Managed APIs (OpenAI, etc.) | Code Frameworks (LangChain, etc.) | |---|---|---|---| | Time to first API | Minutes (one YAML) | Hours (SDK + server code) | Days (framework + integration) | | Provider Coupling | Multi-provider via config | Single provider per SDK | Multi-provider via abstractions | | Code Coupling | Declarative YAML — no application code | Application code required | Framework-specific code required | | Infrastructure Control | Full Sovereignty | Provider-controlled | Heavy Abstraction | | Runtime Flexibility | Hybrid-First (Local + Cloud) | Cloud Only | Complex to customize | | Protocol Support | HTTP / WebSocket / MCP | Provider-specific | Limited | | Data Streaming | First-class across all stages | Response-only (SSE tokens) | Framework-wrapped generators | | Deployment | Docker / Native / Virtualenv / Process | Provider-managed | Manual integration |

Highlights

• Any model, anywhere — HuggingFace, vLLM, llama.cpp locally, or OpenAI/Anthropic/Google/xAI via HTTP
• Agents in YAML — ReAct loops, tool use, multi-step reasoning — no code
• Human-in-the-loop — pause workflows for approval, resume from CLI/UI/API
• 100+ components — models, agents, HTTP/WebSocket clients, vector/graph stores, shell, browsers, and more
• Any protocol — HTTP REST, WebSocket, or MCP with one line
• Any runtime — Docker, native, virtualenv, process, embedded — switch in one line
• Distributed — Redis queue dispatch for horizontal scaling
• Instant Web UI — Gradio-powered UI in 2 lines of YAML
• Streaming everywhere — SSE, WebSocket, and inter-job streams as first-class values

Examples

Browse examples by category:

| Category | What's inside | |---|---| | agents/ | Code reviewer, RAG assistant, Web researcher, Web page analyzer, ... | | showcase/ | End-to-end pipelines: disk analysis, face-based scene search, real-time TTS | | model-providers/ | OpenAI, Anthropic, xAI, Google, ElevenLabs, vLLM | | model-tasks/ | Local chat, embedding, TTS, VLM, face embedding, ... | | media-processing/ | Audio/video/image processing: extraction, conversion, scene detection | | media-broadcast/ | Live broadcasting pipelines (e.g. YouTube Live over RTMP) | | web-automation/ | Web scraping, headless browser automation, page capture | | text-processing/ | Text chunking and preprocessing | | mcp-servers/ | Build MCP servers exposed to Claude, Cursor, ChatGPT | | workflow-queue/ | Redis-backed distributed dispatch (streaming + non-streaming) | | data-streaming/ | Video-to-frames, YouTube live chat, streaming inputs | | job-flow/ | Workflow control: conditional routing, hooks, interrupts | | integrations/ | Vector/graph/KV stores, search engines, channels, tunnels | | runtime/ | Component execution runtimes: Docker, Apple Container, virtualenv, ... |

Browse the full catalog in examples/README.md.

Architecture

Protocol adapters → Composition engine → Runtime executors

!Architecture Diagram

Contributing

We welcome all contributions — bug fixes, docs improvements, new examples.

git clone https://github.com/hanyeol/model-compose.git
cd model-compose
pip install -e .

See CONTRIBUTING if available, or open a PR directly.

License

MIT License © 2025-2026 Hanyeol Cho.

Contact

Have questions, ideas, or feedback? Open an issue or start a discussion on GitHub Discussions.