foxden

An isolated workspace for coding agents on Linux with Docker Engine, natively (e.g. Ubuntu) or in WSL2 on Windows. The agent works in a Linux container with a desktop, browser and normal internet access. It can change project files but can't commit or push.

foxden doesn't ship a coding agent. You add the harnesses, IDEs and services you want as your own Dockerfile layers on top of the base image.

Prototype for local development.

Prerequisites

• Linux with Docker Engine (without Docker Desktop): native, e.g. Ubuntu, or WSL2 on Windows
• a VNC viewer, e.g. TigerVNC
• a Git repository or a folder with several repositories

Details: Prerequisites

Quick start

Put foxden outside your projects:

git clone https://github.com/Foxfire1st/foxden.git \
  ~/.local/share/foxden
S=~/.local/share/foxden/foxden.sh

Start the container and desktop. The first build takes a few minutes:

$S ~/projects/my-repo --desktop

Connect the VNC viewer to localhost:5900. The desktop has a bar with a start menu at the top. On WSL2 you can paste images from the Windows clipboard there with Ctrl+V.

Add your harness as a layer, for example the Cursor CLI (more examples in Layers):

mkdir -p ~/.config/foxden/my-tools
cat > ~/.config/foxden/my-tools/Dockerfile <<'EOF'
ARG BASE=foxden-base:local
FROM ${BASE}
USER root
RUN HOME=/opt/cursor-agent bash -c 'curl -fsSL https://cursor.com/install | bash' \
    && ln -s /opt/cursor-agent/.local/bin/agent /usr/local/bin/agent
USER node
EOF
echo ~/.config/foxden/my-tools > ~/.config/foxden/layers

The next start builds the layer. Sign in once and open the link it shows in your normal browser outside the container (on WSL2: the Windows browser):

$S ~/projects/my-repo env NO_OPEN_BROWSER=1 agent login

Start the agent:

$S ~/projects/my-repo agent

Without a command, $S ~/projects/my-repo opens a shell in the container.

Several repositories together: pass the parent folder instead of the repository, e.g. $S ~/projects/my-product.

Day to day

What	How
Shell in the container	$S ~/projects/my-repo
Harnesses, IDEs, services	your layers, see Layers
End a session	$S ~/projects/my-repo --stop, reports uncommitted changes
Remove the container	$S ~/projects/my-repo --down
Logs of local Docker stacks	foxden-logs.sh follow, see Container logs
Turn on the GPU	echo NVIDIA > ~/.config/foxden/gpu, see GPU

You review, commit and push the agent's changes on the host.

Key rules

1. No Git credentials in the container, not even in an IDE in the container.
2. Don't attach the host IDE to the container with Dev Containers.
3. Review the agent's changes before the host runs them: Git hooks, compose.yaml, build scripts.
4. Only test accounts in the container's browser, no SSO. Open sign-in links in your normal browser.
5. End sessions with --stop and review the changes it reports.
6. Turn off "Send clipboard to server" in TigerVNC.
7. Keep your layers outside every folder you share with the agent.

Reasons: Security model

Documentation

1. Architecture
2. Security model
3. Prerequisites
4. Project files
5. First start
6. Graphical desktop
7. Browser and screenshots
8. Layers: harnesses, IDEs and services
9. GPU
10. Development tools
11. Filesystem and persistence
12. Internet access
13. Managing containers
14. Tests
15. Troubleshooting
16. Roadmap
17. Further reading

Test reports: test-reports/