Docker Agent

Docker Agent is an open-source tool from Docker that runs AI agents defined declaratively in YAML or HCL instead of application code. It runs as a docker CLI plugin (docker agent) and can serve agents through a terminal UI, headless CLI, HTTP API, MCP server, or A2A protocol. The project is licensed under Apache 2.0.

What It Is

Docker Agent is a CLI agent builder and runtime. You describe an agent's model, instructions, toolsets, and sub-agents in a config file, and the runtime handles the LLM orchestration loop, tool execution, multi-agent delegation, and streaming output. The docs state it is not a framework you write code in, not a hosted SaaS, and not a new model.

How the Workflow Runs

You define an agent, start it with docker agent run, and the agent calls tools, reasons, and delegates to sub-agents while results stream back to the TUI or an API client. --exec supports one-shot and piped non-interactive runs, and docker agent serve api exposes an agent as an API. docker agent new generates a config interactively.

Tools, Models, and Teams

Built-in toolsets cover filesystem, shell, git, fetch, memory, RAG, think, todo, and more, and any MCP server (local, remote, or Docker-based) can be added. Multi-agent teams let a coordinator delegate to specialists. Agents can target many providers, including OpenAI, Anthropic, Google Gemini, AWS Bedrock, Mistral, xAI, and local models through Docker Model Runner.

Packaging and Sharing

Agents can be pushed to and pulled from any OCI registry, so docker agent run myorg/agent:tag works like running a container image. The docs also list hooks, permissions, sandbox mode, structured output, model routing, evaluation, skills, and a Go SDK guide.