dwoe: dockerized autonomous coding agents

A simple CLI tool for running autonomous agents in isolated Docker containers.

This is the alpha version software developed for private use! Avoid using it for critical tasks or production setup!

Requirements

• Go 1.25+
• Docker

Quick Start

1. Start Docker daemon
2. Build example images (see /docker folder) or bring a custom one
3. Export CLAUDE_CODE_OAUTH_TOKEN from claude setup-token or set up OPENAI_API_KEY
4. Build the binary and run

make build
./dwoe fire --repo=/path/to/src --do="create a new package to format time"

The fire command launches a container, mounts the input folder, and runs an agentic loop to implement the requirements in the prompt. There is the --work flag to define a prompt file. See other sections for more commands. Check out the examples folder to see more use-cases.

Usage

There are multiple commands to run and inspect workers.

dwoe <version>

Usage:
	dwoe [flags] <command> [args]

Flags:
	--datadir <dir>     Data directory (default: ~/.dwoe)
	--logfile <path>    Write JSON logs to file
	--loglevel <level>  Log level: debug, info, warn, error (default: warn)
	--logfmt <format>   Log format: text (human), json (default)
	--noproxy           Disable proxy container

Commands:
  batch    <dir>                      Run all task files in parallel
  collect  <name|id> [--batch ID]     Collect commits into a repo branch
  create   <task.yaml>                Create workspace from config
  destroy  <name|id>                  Remove workspace
  fire     --repo <url|path> [flags]  Quick-start workspace from repo
  inspect  <name|id>                  Show detailed workspace info
  list     [--format FMT]             List workspaces
  logs     <name|id>                  Show workspace logs
  patches  <name|id> --dir <dir>      Export patches to directory
  run      <task.yaml>                Create and start workspace
  start    <name|id>                  Start existing workspace
  status   <name|id>                  Show workspace status
  stop     <name|id>                  Stop running workspace
  version                             Show version
  web      [--addr ADDR]              Start web dashboard

The default location for workspace data, including logs, code artifacts, and task config, is ~/.dwoe.

Use dwoe list to show the list of running and finished workers.

Use dwoe inspect <name|id> to get information about the worker.

Each dwoe fire|run starts two containers:

• The one with agentic loop implementing the task, and
• The proxy container that prevents arbitrary web requests.

There are a few Docker files to build default agent containers. A custom container can be used instead. See examples and default config. Use YAML files and run command to redefine the image, or build the example images with make images.

Every agent image carries both agent CLIs, so the same image runs any provider: dwoe-agent:latest is the universal default, with smaller dwoe-agent:go|python|c|cpp variants per language. See docker/README.md.

Providers

The agent backend is selected with agent.provider in the task file, or --provider on dwoe fire. Supported values are anthropic (default) and openai.

Authentication

Credentials are passed through from the host environment; nothing is stored by dwoe.

Provider	Environment variable
anthropic	CLAUDE_CODE_OAUTH_TOKEN or ANTHROPIC_API_KEY
openai	CODEX_API_KEY or OPENAI_API_KEY

Build a custom image to modify the entry point and provide the authentication credentials differently.

Task file

Running a worker from a task file:

dwoe run task.yaml

The file format:

name: my-task
source:
  local_path: ./repo
  prompt_file: ./prompt.md
agent:
  provider: anthropic
  image: dwoe-agent:go
  model: claude-sonnet-5
  max_turns: 10
  env_vars:
    CLAUDE_CODE_OAUTH_TOKEN: ${CLAUDE_CODE_OAUTH_TOKEN}
git:
  user_name: "dwoe-agent"
  user_email: "agent@dwoe.dev"
resources:
  cpu: "2"
  memory: "4G"

Web Dashboard

Use dwoe web to start a dashboard for monitoring workers, streaming logs, and viewing workspace status in the browser.

dwoe web --addr :9090

Examples

See examples/ for complete walkthroughs covering quick start, single tasks, Python projects, batch parallel runs, patch collection, batch-and-merge workflows, and custom prompt configuration.

Development

make test
make lint
make all

Note that the project includes .air.toml and make dev allows hot reloading upon changes.