👋 Welcome to the Docker Sandboxes with GitHub Copilot Lab!
In this hands-on lab you will use GitHub Copilot inside a Docker Sandbox to work on DevBoard, a small issue tracker (FastAPI + Next.js) located in the project/ folder. Along the way you will see how sandboxes keep your machine, your files and your credentials safe while an AI agent does real work.
Duration: about 100 minutes, plus the prerequisites, which you do beforehand.
- Control what an agent can reach on the network with network policies
- Understand how kits define an agent's sandbox, and build your own
- Give an agent access to services without ever exposing your credentials
- Run an agent in direct mode and in isolated clone mode, and run several agents in parallel
- Monitor sandboxes from the TUI dashboard
- Run Docker inside a sandbox and reach it from your browser with port forwarding
- Give your agent tools through the sandbox's MCP gateway
A Docker Sandbox is not a container: it's a lightweight microVM, with its own kernel, that runs one AI agent. Everything the agent can touch on the way out goes through a few doors that you control from your machine.
graph LR
You(["You, in your browser"])
subgraph Host["Your machine"]
direction TB
Folder["Your project folder"]
Keychain["OS keychain<br/>(your tokens)"]
Proxy["Proxy<br/>network policy + credentials"]
Gateway["MCP gateway"]
end
subgraph Sandbox["Sandbox (microVM)"]
direction TB
Agent["Agent (Copilot)"]
Docker["Private Docker daemon<br/>and containers"]
end
subgraph Outside["Outside world"]
direction TB
Internet["Allowed hosts<br/>(GitHub, PyPI...)"]
McpServers["MCP servers"]
end
Folder <-->|workspace mount| Agent
Agent -->|every request| Proxy
Keychain -->|token added on the way out| Proxy
Proxy -->|allowed hosts only| Internet
Agent -->|tool calls| Gateway
Gateway -->|calls| McpServers
Agent -->|runs| Docker
You -->|sbx port| Docker
%% Styles — containers
style Host fill:#E5F2FC,stroke:#1D63ED,color:#00084D
style Sandbox fill:#f0faf4,stroke:#86c5a0,color:#00084D
style Outside fill:#fff7f0,stroke:#d4915e,color:#00084D
%% Styles — host nodes
style Folder fill:#fff,stroke:#0DB7ED,color:#384D54
style Keychain fill:#fff,stroke:#0DB7ED,color:#384D54
style Proxy fill:#E5F2FC,stroke:#1D63ED,color:#00084D
style Gateway fill:#E5F2FC,stroke:#1D63ED,color:#00084D
%% Styles — sandbox nodes
style Agent fill:#e0f2e6,stroke:#6db88a,color:#384D54
style Docker fill:#e0f2e6,stroke:#6db88a,color:#384D54
%% Styles — external nodes
style Internet fill:#fff0e5,stroke:#d4915e,color:#384D54
style McpServers fill:#fff0e5,stroke:#d4915e,color:#384D54
%% Styles — actor
style You fill:#fff,stroke:#1D63ED,color:#00084D
| Door | What it means | Module |
|---|---|---|
| Network proxy | Every outbound request leaves through a proxy on your machine, which allows or blocks it by host | 01 |
| Kit | A YAML file that says which image the sandbox runs, how the agent starts and what it's given | 02 |
| Credential injection | Your tokens stay in your OS keychain; the proxy adds them to the agent's requests, so the agent only ever sees a placeholder | 03, 04 |
| Workspace | Only the folder you give the sandbox is shared with it, not the rest of your disk; or the agent gets a private Git clone instead | 04, 07 |
| Private Docker daemon | Containers the agent starts live inside the microVM, invisible to your host's Docker | 08 |
| Port publishing | Nothing inside is reachable from your machine until you publish a port | 09 |
| MCP gateway | Tools reach the agent through the gateway, without opening its network. The MCP servers run outside the sandbox, so its network policy doesn't apply to them: register only servers you trust | 10 |
Because the agent can't step outside these doors, you can let it work without approving every command (--yolo), and still know what it can and can't reach.
| # | Module | You will... | Time |
|---|---|---|---|
| 00 | Prerequisites | Check your accounts and tools | before the lab |
| 01 | Network policies | Start your first sandbox, lock the network down, then allow what you need | 10 min |
| 02 | Kits | See how an agent's sandbox is defined, and build a kit | 8 min |
| 03 | Secrets and credentials | Store a token the agent can use but never read | 3 min |
| 04 | Agent sandbox | Create your agent's sandbox, check its isolation, and meet your agent | 12 min |
| 05 | TUI dashboard | Monitor sandboxes and network activity | 3 min |
| 06 | Run in direct mode | Let the agent fix failing tests in your working tree | 12 min |
| 07 | Clone mode and parallel agents | Run isolated agents side by side and merge their work | 20 min |
| 08 | Docker inside the sandbox | Start the full app with Docker Compose in a sandbox | 8 min |
| 09 | Port forwarding | Open the app running in the sandbox in your browser | 5 min |
| 10 | MCP servers | Give your agent a tool without opening its network, then scan the images it suggests | 12 min |
| 11 | Wrap up | Clean up and find out where to go next | 3 min |
Stuck? See Troubleshooting.
-
Clone the repository:
git clone https://github.com/mathieu-benoit/sbx-lab cd sbx-labYou commit the agents' work locally during the lab, but you never push it, so you don't need a fork.
-
Complete the prerequisites before the lab: they download several hundred megabytes, which is slow on shared Wi-Fi.
-
Follow the modules in order.
Enjoy!