Docker Compose Tries to Tame the Agent Framework Zoo

What it does

This repo is essentially a curated recipe book of Docker Compose files that package popular AI agent frameworks—CrewAI, LangGraph, Google ADK, Agno, and others—alongside local LLM inference via Docker Model Runner and various MCP tool servers. Each demo is self-contained, aiming to let you spin up a multi-agent fact checker, a SQL agent, or a travel planner without installing a dozen Python environments on your host. It handles the plumbing so you can compare how different frameworks talk to models and tools.

The interesting bit

The clever part isn’t any single agent, but the normalization layer: Docker is positioning Compose as the universal orchestrator for the fragmented agent ecosystem, treating LLMs, vector stores, and MCP servers as just more containerized dependencies. It also nudges you toward Docker Model Runner for local inference, though most demos graciously include an OpenAI escape hatch if your laptop lacks the VRAM.

Key highlights

• Covers 12 demos spanning frameworks like CrewAI, LangGraph, Google ADK, Vercel AI SDK, Spring AI, and MinionS.
• Supports local model inference via Docker Model Runner (Qwen3, Llama 3.2, Gemma 3) or remote APIs like OpenAI and Cerebras.
• Integrates MCP servers—DuckDuckGo, Brave, GitHub, Postgres, Wikipedia, and more—as composable sidecars.
• Self-contained stacks meant to run locally or against a cloud Docker context.
• Dual-licensed under Apache 2.0 or MIT.

Caveats

• Requires a GPU or Docker Offload for local models; the README notes specific driver requirements that can trip up Linux and Windows users.
• Most demos need manual environment file setup from provided examples, so the out-of-the-box experience is not quite zero-config.
• The repo is examples-only; it is not a new framework or runtime itself, just integration glue between existing projects.

Verdict

Worth a look if you are evaluating agent frameworks and want a clean-slate way to test them without polluting your machine. Skip it if you are already committed to one stack and comfortable managing your own dependencies; this is a tasting menu, not a production platform.