Coder - Cloud Development Environments

Coder is a self-hosted platform for deploying cloud development environments. It enables LBNL researchers to run VS Code, JetBrains IDEs, and AI coding agents in secure, network-isolated containers on Berkelium’s Kubernetes infrastructure.

Why Use Coder on Berkelium?

Self-Hosted Security: Your code and data never leave the Berkelium cluster — critical for compliance-sensitive research
AI-Native Development: Run AI coding agents (Cursor, Devin Desktop, GitHub Copilot) with centralized governance and audit trails
Consistent Environments: Template-based workspaces ensure reproducible development setups across your research team
GPU Acceleration: Access NVIDIA H100, A100, and L40S GPUs directly from your cloud IDE
Network Isolation: Enforce zero-trust networking policies between workspaces and external resources

Quick Start

1. Request a Coder Workspace

Visit the Coder Portal and authenticate with your LBNL OneID.

2. Choose a Workspace Template

Coder offers pre-configured templates for common research workflows:

Template	CPU	GPU	Memory	Use Case
Standard Dev	4 cores	None	16GB	General Python/R development
Data Science GPU	8 cores	A100 (40GB)	64GB	ML model training, JupyterLab
LLM Research	16 cores	H100 (80GB)	128GB	Large language model fine-tuning
Vision Workloads	12 cores	L40S (48GB)	96GB	Computer vision, image processing

3. Connect Your IDE

Once your workspace is running:

VS Code (Recommended):

# Install Coder CLI
curl -L https://coder.com/install.sh | sh

# Connect to your workspace
coder code-server <workspace-name>

JetBrains IDEs:

Open your JetBrains IDE (IntelliJ, PyCharm, CLion)
Install the Coder Gateway plugin
Connect to https://coder.berkelium.lbl.gov
Select your workspace and launch

Browser-Based:

Access VS Code directly at https://coder.berkelium.lbl.gov/workspace/<name>
Full terminal, file explorer, and extension support

Advanced Usage

Custom Workspace Templates

Create custom templates for your research group by defining a .coder/coder.yaml file in your repository:

workspace:
  name: my-research-env
  image: ghcr.io/my-lab/custom-ml-image:latest
  resources:
    cpu: 16
    memory: 64Gi
    gpu: nvidia.com/gpu=1
  mounts:
    - source: pvc/my-datasets
      destination: /home/coder/datasets
      read_only: false
  ports:
    - port: 8888  # JupyterLab
      expose: true
    - port: 6006  # TensorBoard
      expose: true

AI Agent Governance

Coder provides centralized control over AI coding assistants:

Audit Logs: Track all AI tool invocations and code suggestions
Policy Enforcement: Block or allow specific AI models based on data sensitivity
Network Controls: Restrict AI agents from accessing external APIs or datasets
Cost Monitoring: Monitor token usage and API costs across your team

Persistent Storage

Mount CephFS persistent volumes to retain your work across sessions:

# In your workspace terminal
kubectl patch workspace <workspace-name> -n sci-myproject --type='json' \
  -p='[{"op": "add", "path": "/spec/volumes/-", "value": {
    "name": "home-storage",
    "persistentVolumeClaim": {"claimName": "my-home-pvc"}
  }}]'

Best Practices

Stop Idle Workspaces: Coder workspaces consume GPU/CPU resources even when idle. Stop them when not in use:
Terminal window
```
coder stop <workspace-name>
```
Use Templates: Standardize environments across your team using workspace templates for reproducibility.
Mount Persistent Storage: Always mount a PVC for your home directory and datasets to avoid data loss.
Leverage AI Governance: For sensitive research code, enable AI governance policies to control model access.
Port Forwarding: For web applications running in your workspace:
Terminal window
```
coder port-forward <workspace-name> 8080:8080
```

Troubleshooting

Workspace Won’t Start

Check resource quotas: kubectl describe quota -n sci-myproject
Verify GPU availability: kubectl describe nodes | grep -A 5 "Allocated resources"
Review workspace logs: coder logs <workspace-name>

IDE Connection Issues

Ensure your Coder CLI is up to date: coder version
Check network connectivity: coder ping <workspace-name>
Restart the code-server: coder restart <workspace-name>

GPU Not Detected

Verify GPU resource requests in your template
Check NVIDIA device plugin: kubectl get pods -n kube-system | grep nvidia
Restart workspace with GPU flag: coder start <workspace-name> --gpu

Resources

Coder Documentation — Official guides and API reference
Coder Portal — Launch and manage workspaces
Workspace Templates Repository — Community-contributed templates
Slack: #coder-users — Get help from other researchers

Need Help?

Office Hours: Tuesday & Thursday, 11am – 12pm PT (Zoom link in Slack)
Slack: #coder-users on the LBNL workspace
Help Desk: Submit a ticket for Coder-specific issues