coding-agent-box

GPU-enabled Docker environment for coding agents on a fresh Linux machine.

Prerequisites

• Docker Desktop or Docker Engine
• Docker Compose v2 (docker compose)

Included tools

• Node.js
• pixi
• git
• GitHub CLI (gh)

Included coding agents

• OpenCode
• Claude Code
• OpenAI Codex CLI
• PI Coding Agent

Host directory layout

Create these directories on the host before starting the container:

mkdir -p ~/coding-agent-box/projects
mkdir -p ~/coding-agent-box/configs/.opencode
mkdir -p ~/coding-agent-box/configs/.claude
mkdir -p ~/coding-agent-box/configs/.codex
mkdir -p ~/coding-agent-box/configs/.pi

Start

Pull and launch the published image:

docker compose pull
docker compose up -d

To pin a specific published version:

CODING_AGENT_BOX_IMAGE=ghcr.io/computbiol/coding-agent-box:v0.1.0 docker compose up -d

Enter the container:

docker exec -it coding-agent-box bash

Local build

To build the image locally instead of pulling from GHCR:

docker compose -f docker-compose.yml -f docker-compose.build.yml up -d --build

Maintenance

• Release instructions for maintainers are in docs/releasing.md.

Notes

• GPU support requires the host machine to have NVIDIA Container Toolkit installed.
• The container runs as user agent and the user has passwordless sudo access.
• Project files are mounted from ~/coding-agent-box/projects to /home/agent/projects.
• Agent config directories are mounted from ~/coding-agent-box/configs/* to /home/agent/.*.