Docker released Docker Agent, a CLI plugin for AI agents in YAML

An AI agent in production usually travels on a separate stack from the services it talks to. Docker released Docker Agent this week to collapse that: define the agent in YAML, push it to any OCI registry, run it with the same Docker CLI you already use. The plugin is Apache 2.0 and ships for Docker Desktop and standalone systems.

Docker says docker-agent "lets you create and run intelligent AI agents that collaborate to solve complex problems, no code required". The command runs under docker agent when installed as a plugin, or as docker-agent on its own. The project is on GitHub at docker/docker-agent with 3,477 stars and is on version v1.149.0, with the last five releases landing in a twelve-day window.

What the YAML file does

An agent is a block of YAML that names a model, an instruction, and a list of toolsets. The example in the README picks openai/gpt-5-mini, writes two lines of system prompt, and attaches a DuckDuckGo tool through MCP. One command, docker agent run agent.yaml, starts it. A single file can define several agents and let them delegate to each other, which Docker calls multi-agent architecture.

Docker says the plugin is "AI provider agnostic". The list in the README names OpenAI, Anthropic, Gemini, AWS Bedrock, Mistral, xAI and Docker Model Runner, the local-model runtime Docker ships. One API key is enough to start.

Tools, retrieval and MCP

Four built-in tool types ship with the plugin: think, todo, memory, and retrieval. The retrieval layer is pluggable and the README lists BM25, embeddings, hybrid search and reranking as supported out of the box. Any MCP server can be attached as a toolset, local or remote, including Docker's own catalogue reached through docker: references.

Shipped the same way as a container

The piece a team already using Docker will recognise: an agent can be pushed to any OCI registry and pulled from one. docker agent run myorg/agent:tag reads the agent definition from a registry the same way docker run reads an image, so the distribution path is one the engineering team is already using for its services. Docker Desktop 4.63 and newer bundle the plugin; Homebrew and GitHub Releases are the other install paths.

The project supports macOS and Linux. Windows users route through WSL. A GitHub Action and MCP mode let the same agent run inside a CI pipeline or be called by another agent.

Docker's own build loop uses the plugin: the contributing guide says docker agent run ./golang_developer.yaml is how the Docker Agent team builds Docker Agent.

For a team that already runs Docker for every service, keeping its agents in the same registry, with the same CLI and the same build and distribution path, removes a parallel set of tools from the mental overhead of running AI features in production.

Source

• docker/docker-agent on GitHub (the project README)
• Docker Agent documentation
• Hacker News discussion