Basic usage¶

First install the plugin with the two commands on the home page, then type /lab-setup once in Claude Code. From then on the skills keep themselves up to date, in Claude Code, Codex and Kimi Code alike.

Talk to your agent about cluster work and the right skill loads itself. In Claude Code you can also call one directly: /lab-hpc, or lab-jupyter, or lab-containers.

lab-edison is the exception. It never loads on its own, because every run spends your own credits. Ask for it by name.

Things you can ask¶

"Test this repo on chimera." The agent moves your code the lab way, local to GitHub to a git pull on the cluster. It takes a compute node, never the login node, and runs the tests there. You see the exact sbatch or srun first.

"Get me a GPU node on chimera." It drafts a job on the lab's free zhoulab_gpu_priority partition and asks what you want: CPUs, memory, GPUs, time.

"Start Jupyter Lab on the cluster and open it on my laptop." lab-jupyter checks whether you already have a Jupyter job running, and reuses it. Otherwise it submits one with your approval, opens an ssh tunnel, and hands you a http://localhost:9990/lab?token=… link that works in your browser. Either cluster.

"Update the ml container image on ircbc." lab-containers compares your image's recorded digest against the registry and tells you whether an update is actually needed. Only then does it pull and rebuild, pulling on the login node and building in a job, because ircbc's compute nodes have no internet.

"Run this alignment on ircbc." The command runs inside the right lab container, in a Slurm job, with the environment active. The bare OS there is far too old for modern tools.

"/lab-edison what is known about …" lab-edison checks your key file, names the job it picked, prints your question in full, and only then sends it. You get an answer with citations, and the task id that paid for it. No cluster on this path.

For a goal too big for one question, ask about Kosmos. The skill drafts the goal with you and checks your data for free, and starts a run only when you ask for that run.

This is the one skill that wants a tool of its own: pixi on your PATH. Your first Edison command then builds what it runs in before it sends anything, so it can look hung. It is not. The build runs once more after each plugin update. See its page.

What the agent will ask you¶

Which cluster, unless your request already makes it obvious. What the job needs: CPUs, memory, GPUs, wall time. And your confirmation of the exact command before any job is submitted.

It also asks before cancelling or deleting anything shared. Reservation jobs and container images belong to the whole group.

Tips¶

Fill in ~/.claude/lab/personal.md with your Jupyter port, launch path and usual job resources, and the agent stops re-asking session after session.

Name the cluster in your request ("…on chimera", "…on ircbc") to skip the which-cluster question.

The skills tell the agent to read your ssh config, never to change it. SSH setup stays in your hands.