Docker Agent is an AI agent builder الطبقات released by Docker Engineering that enables users to create intelligent agents through a simple YAML file. The configuration declares a model (such as openai/gpt-5-mini), provides a description and instruction, and assigns toolsets—each toolset can reference an MCP server like docker:duckduckgo. This declarative approach makes agents versionable and shareable, and the framework supports multi‑agent orchestration where teams of specialized agents delegate tasks automatically.

GitHub repository preview for docker/docker-agent showing the AI Agent Builder and Runtime project by Docker Engineering
Docker Agent project repository showing the official AI agent framework released by Docker Engineeringgithub.com

This repository image represents the Docker Agent project introduced in the article's opening as an AI agent builder released by Docker Engineering

To get started, users can install the docker-agent CLI plugin via Docker Desktop (version 4.63+ where it is pre‑installed), Homebrew (brew install docker-agent), or by downloading a binary release from GitHub and symlinking it to ~/.docker/cli-plugins/docker-agent. After setting an API key for a cloud model or using Docker Model Runner for local models, they can run the default agent with docker agent run, generate a starter configuration interactively via docker agent new, execute a custom YAML file with docker agent run agent.yaml, or launch an agent directly from an OCI registry using docker agent run myorg/agent:tag.

Animated demonstration showing the Docker Agent CLI in action with various commands like docker agent run and docker agent new
Docker Agent CLI demonstration showing how to run agents, create configurations, and deploy from registriesgithub.com

This animated demo illustrates the practical usage of Docker Agent commands discussed in the getting started section, including running agents, generating starter configurations, and deploying from OCI registries

Docker Agent’s rich tool ecosystem includes built‑in tools plus any MCP server (local, remote, or Docker‑based), and it is AI provider agnostic, working with OpenAI, Anthropic, Gemini, AWS Bedrock, Mistral, xAI, Docker Model Runner and others. Advanced reasoning capabilities such as think, todo, and memory tools are provided, along with pluggable RAG support that can use BM25, embeddings, hybrid search and reranking. Once defined, agents can be packaged as OCI images and pushed to any registry for distribution and reuse.

For isolation, Docker Agent offers a sandbox mode that launches agents inside a Docker sandbox VM. When the --sandbox flag is used, Docker Agent delegates to the standalone sbx CLI if available, otherwise falling back to the docker sandbox CLI plugin. The agent’s working directory is mounted read‑write inside the VM, while the configuration directory and staged kit are mounted read‑only, preventing guest aliases from altering the selected agent. Cloud sandboxes provide a fresh remote workspace and do not upload the host’s directory, local YAML, user config or auto‑kit by default; any needed files or secrets must be supplied through a workload kit, and sandbox IDs and cleanup hints are printed to stderr on exit.