Vibeleaderboard
Index / tool
Visit github.com
Category
AI Agents
Rank
Pricing
Open Source
Type
TOOL
Use case
Agent Building
Latest release
v1.163.0
Date

About

Docker Agent is an AI agent builder and runtime from Docker Engineering. It lets users build, run, and share agents using a declarative YAML config, a tool ecosystem, and multi-agent orchestration, without writing code.

What it does

A command-line program, shipped as a plugin for the docker command, that reads a short text file naming a model, a system prompt and a list of capabilities (file access, shell, web search, memory), then drives the loop of calling the model and executing what it asks for. Agents can hand work to other agents, either as a parent waiting on a specialist or by passing the whole conversation along. Finished agents are pushed to any container registry and pulled by name, the same way images are.

Why it's ranked here

The case rests on scope and provenance. It comes from Docker, is written in Go, and ships pre-installed in Docker Desktop as well as through Homebrew and prebuilt binaries. One definition can be served as a terminal UI, a headless command, an HTTP API, an OpenAI-compatible chat endpoint, an MCP server or over the A2A agent protocol. It talks to OpenAI, Anthropic, Gemini, Bedrock, Mistral, xAI and local models, and the documentation treats safety and spending as first-class settings rather than afterthoughts.

What's good

The guardrails are specific. Permissions use deny, allow and ask rules with argument matching, and deny always wins, even over an agent's own allow. Four named safety modes decide what happens to calls no rule covers, including a fail-closed mode for unattended runs. Spending caps cover dollars, tokens and working time; agents that name the same cap share one pot, so fanning out to sub-agents cannot multiply the limit. Sandbox mode runs the agent inside a Docker sandbox VM with a declared network allowlist. Shared agents pinned to a digest load from cache with no network lookup.

Tradeoffs

The documentation names several limits itself. Spending caps are checked between turns, so a run can overshoot by the turn in flight, and counters live in memory, so restarting the API server resets them. Models without pricing data add nothing to the dollar cap unless priced by hand. The restricted safety mode is described as defense in depth, not a security boundary. A config pulled from a registry can declare autonomous mode, which runs every tool call unprompted unless your own setting overrides it. Anonymous usage data is collected by default.

How to use it well

Fits a developer who wants a coding, ops or research agent in a terminal or pipeline without writing the model loop by hand. Set your own safety floor in user settings before running anyone else's agent, since that outranks the author's default. Pin external sub-agents to a digest for fast, reproducible startup. Put a budget with some headroom on any unattended job, and use the sandbox when shell access matters. It is not a coding framework, a hosted service or a model, so teams needing custom logic in Python or TypeScript should look elsewhere.

Technical notes+

The module in go.mod is github.com/docker/docker-agent; main.go hands argv to cmd/root Execute and maps cli.StatusError to the exit code. Dependencies include the official Anthropic, OpenAI and Google genai SDKs, the MCP go-sdk (replaced with an unreleased commit pending a tagged fix), a2a-go, Google ADK, bubbletea for the TUI and modernc sqlite for session storage. The Dockerfile cross-compiles static CGO binaries for Linux, macOS and Windows with xx and zig, and builds a docker/sandboxes template image. pkg/app/app.go wires runtime events into the TUI through a buffered channel throttled at 50ms. docs/configuration/budget/index.md and docs/configuration/permissions/index.md document the cap and permission semantics.

Observed

Language
Go
Interfaces
Terminal UI, headless CLI, HTTP API server, OpenAI-compatible chat endpoint, MCP server, A2A protocol
Install surface
Docker CLI plugin pre-installed in Docker Desktop; Homebrew; prebuilt binaries from GitHub Releases; container image
Platform support
Static binaries cross-built for Linux, macOS and Windows
Agent definition format
YAML or HCL files
Distribution
Agents pushed to and pulled from any OCI registry
Model providers
OpenAI, Anthropic, Gemini, AWS Bedrock, Mistral, xAI, Docker Model Runner and others
Tool extension
Built-in toolsets plus MCP servers over Docker MCP Gateway, stdio or remote HTTP
Telemetry
Anonymous usage data collected, per the README
License
Not stated in the files read

Read from README.md, go.mod, main.go, Dockerfile, docs/index.md, docs/concepts/agents/index.md, docs/concepts/tools/index.md, docs/configuration/permissions/index.md, docs/configuration/sandbox/index.md, docs/configuration/budget/index.md, docs/concepts/multi-agent/index.md, pkg/app/app.go.

Tags

ai-agentsdockeryamlorchestrationruntimeai

Tech Stack

GoDocker

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.