How to Set Up Claude Code on the Cluster (CCCCC)¶

What is Claude Code?¶

Claude Code is Anthropic's command-line coding agent. It runs in your terminal and can read files, write code, and execute commands on your behalf so you can spend your attention where it matters. On this cluster, we provide a shared configuration that teaches Claude the rules of our environment — Slurm, shared filesystems, GPU etiquette, and common pitfalls — so it behaves as a responsible cluster citizen from the first session.

Prerequisites¶

• An active Claude Pro or Max subscription, or an Anthropic API key
• SSH access to the cluster

Step 1: Install Claude Code¶

curl -fsSL https://claude.ai/install.sh | bash

Verify it installed:

claude --version

Step 2: Go to Your User Directory¶

Always start from your workspace on the shared filesystem:

cd /data/$USER

This is where all your projects, conda environments, datasets, and outputs should live. Never work out of $HOME for anything large — home directory quotas are small.

You may want to add this to your ~/.bashrc so you land here on every login:

echo 'cd /data/$USER' >> ~/.bashrc

Step 3: Authenticate¶

Run claude once to trigger the login flow. It will open a browser URL — if you're connected via SSH, copy the URL and open it on your local machine to complete authentication.

If you're using an API key instead of a subscription:

export ANTHROPIC_API_KEY="sk-ant-..."

Add that export to your ~/.bashrc to persist it across sessions.

Step 4: Set Up the Cluster Protocol¶

We maintain a shared file at /data/claude/cluster-citizen.md that teaches Claude how to behave on this cluster. To load it automatically every time you run Claude, add this alias to your ~/.bashrc:

echo 'alias claude="claude --append-system-prompt-file /data/claude/cluster-citizen.md"' >> ~/.bashrc
source ~/.bashrc

That's it. Now every time you type claude, the cluster rules are injected alongside Claude's default instructions and your personal configuration. Your ~/.claude/CLAUDE.md and any project-level settings continue to work normally.

Verify it's working¶

Start a session and ask:

> What partitions are available on this cluster?

Claude should respond with cais, cais_cpu, tamper_resistance, and tamper_resistance_cpu without you needing to tell it.

Step 5 (Optional): Add Your Own CLAUDE.md¶

The cluster protocol covers shared rules. For your own project-specific context, create a CLAUDE.md in your project directory:

cd /data/$USER/my-project
cat > CLAUDE.md << 'EOF'
# My Project

- Python 3.11, PyTorch 2.x
- Conda env: myenv
- Tests: pytest tests/
- Data lives in /data/$USER/datasets/imagenet/
EOF

Claude loads this automatically when you run claude from that directory. It stacks with the cluster protocol — both apply.

Usage Tips¶

Use tmux for session persistence¶

You're SSHing into the cluster. If your connection drops, your Claude Code session dies — unless you're inside tmux. Always start Claude inside a tmux session:

# Start a new tmux session
tmux new -s claude

# Now start Claude as normal
claude

# If your SSH drops, reconnect and reattach:
tmux attach -t claude

Your conversation, running processes, and Claude's context are all preserved.

tmux is also required if you want to use agent teams — Claude Code's feature for spawning multiple agents working in parallel in separate panes. Without tmux, agent teams still work but only in a single-terminal mode where you cycle between agents with Shift+Up/Down.

Quick tmux reference:

Always use Slurm for heavy work¶

Claude knows this rule, but reinforce it by telling Claude what you're doing:

> I need to train a model on 2 GPUs for ~6 hours. Write me an sbatch script.

Claude will use the correct partition names and resource flags for this cluster.

Use conda, not pip --user¶

All Python environments should be conda envs stored under /data/$USER/. If you don't have miniconda yet, follow the How to Install Miniconda or Anaconda guide.

Reporting mistakes¶

If Claude does something wrong that other users should be protected from — like trying to run training on the login node, or installing packages to the wrong location — report it to Jason Lim via Slack. We maintain a shared lessons file that gets updated so every Claude session on the cluster learns from past mistakes.

Quick Reference¶

Troubleshooting¶

"command not found: claude"¶

The install script may not have added Claude to your PATH. Try opening a new shell, or run source ~/.bashrc. If it still fails, check that ~/.claude/bin (or wherever the installer placed it) is in your PATH.

Claude doesn't know about the cluster¶

The alias isn't set. Run alias claude to check. If it doesn't show the --append-system-prompt-file flag, re-run the Step 3 command.

Authentication fails over SSH¶

Copy the URL Claude prints and open it in a browser on your local machine. Complete the login there. The CLI will detect the auth automatically.

Claude tries to run training on the login node¶

Tell it explicitly: "This must go through Slurm." If the cluster protocol is loaded, Claude should already know this — if it doesn't, check that your alias is set correctly.

Home directory is full¶

Likely caused by pip/conda installing to $HOME. Move your conda installation to /data/$USER/miniconda3 and set HF_HOME as described above.

Contact the administrator if none of the solutions are satisfactory.