Multi-Agent Dev Environment

A single Docker Compose workspace for Claude Code, GitHub Copilot CLI, and OpenAI Codex, all sharing a common base image and the same workspace/ directory.

---

Repository Structure

.
├── Dockerfile.base      # Shared base: Ubuntu 24.04 + Node.js + Python 3 + Rust + bash
├── Dockerfile.claude    # FROM dev-base + Claude Code CLI
├── Dockerfile.copilot   # FROM dev-base + GitHub Copilot CLI
├── Dockerfile.codex     # FROM dev-base + OpenAI Codex CLI
├── docker-compose.yml   # All three agents + PostgreSQL
├── .gitignore
├── workspace/           # Shared project files — all agents mount this
├── claude-config/       # Persistent Claude Code auth - gitignored so no secrets can leak
├── copilot-config/      # Persistent Copilot CLI auth - gitignored so no secrets can leak
├── codex-config/        # Persistent Codex config - gitignored so no secrets can leak
├── commandhistory/      # Persistent command history - gitignored so no data can leak
│   ├── claude/
│   ├── copilot/
│   └── codex/
└── README.md

---

How the Base Image Works

Dockerfile.base builds a shared image called dev-base containing everything all three agents need: system packages, Node.js LTS, Python 3, Rust, pnpm, and yarn. Each agent Dockerfile starts with FROM dev-base and only adds its own CLI tool.

Benefits:

• Common layers are built and cached once — not three times
• Rebuilding one agent image is fast (only its own layer changes)
• All agents run identical Node/Python/Rust versions
• Disk space: shared layers are stored once by Docker

The base image uses plain bash with Ubuntu's default color prompt, tab completion (bash-completion), and persistent history — no heavy shell frameworks needed.

---

Quick Start

1. Create host directories

mkdir -p commandhistory/claude commandhistory/copilot commandhistory/codex

2. Build the base image first

docker build -f Dockerfile.base -t dev-base .

3. Build all agent images and start the containers

docker compose up --build -d

4. Open an interactive shell in any agent

docker compose exec -it claude-code bash
docker compose exec -it copilot bash
docker compose exec -it codex bash

5. Authenticate each agent (first time only)

All three agents use an interactive device code flow — no API keys or .env files needed. Auth is persisted in the *-config/ directories on the host and survive rebuilds.

Claude Code (inside claude-code container):

claude
# follow the interactive login flow

Copilot CLI (inside copilot-dev container):

copilot
# follow the interactive login flow

Codex (inside codex-dev container):

codex
# follow the interactive login flow

---

Working Across Agents

All three containers mount the same ./workspace directory.

You can use the agents to work on different projects or together on one project.

Use Git as your coordination layer — commit before switching agents so you always have a clean handoff point:

# for example in claude-code: 
# checkout a new feature branch
git checkout -b feature-branch

# finish your task with claude code, review it, then commit
git add . && git commit -m "claude: add new feature"

# then merge the feature into the main branch
git checkout main
git merge feature-branch

A suggested task split based on each tool's strengths:

Agent	Best for
Claude Code	Architecture, refactoring, tests, large codebase understanding
Copilot CLI	GitHub integration (issues, PRs, branches), multi-model tasks
Codex	Focused code generation, OpenAI model preference

---

Security

Each agent runs in its own isolated container with separate memory and environment variables. Auth tokens and config are mounted into only the container that needs them — the Copilot config is invisible to the Claude and Codex containers, and vice versa.

The only shared surface between containers is the workspace/ directory (intentional) and the host network (localhost). No MCP servers, web search, or external tool access is configured by default, keeping the attack surface minimal for local development.

---

Useful Commands

# Start all containers
docker compose up -d

# Start only one agent
docker compose up -d claude-code

# Open a shell
docker compose exec -it claude-code bash

# Stop everything
docker compose stop

# Stop and remove containers
docker compose down

# Rebuild base image after changes
docker build -f Dockerfile.base -t dev-base .
docker compose up --build -d

# Remove everything including volumes (⚠️ deletes PostgreSQL data)
docker compose down -v

---

PostgreSQL

Available to all agents via localhost:5432 (host network mode).

Setting	Value
Database	quest_db
User	quest_user
Password	quest_passw0rd

Change the default credentials before using this in any shared or production environment!

---

License

MIT License

Copyright (c) 2026 Michael Schaefer https://github.com/mischa-robots/multi-agent-docker