Using Claude Science with Lawrencium¶

Claude Science is a desktop application from Anthropic that pairs Claude with an analysis environment on your own computer. You describe a research task in plain language, and Claude writes and runs Python, R, or shell code, reads the folders you grant it access to, and saves results as versioned artifacts with a provenance record.

Through its Remote compute clusters feature, Claude Science can also run jobs on a Slurm cluster that you reach over SSH. This page explains how to add Lawrencium as a remote compute cluster so that Claude can submit Slurm jobs to the Lawrencium and Einsteinium clusters on your behalf.

Claude Science documentation

This page only covers the Lawrencium-specific setup. For installation, plans, usage limits, permissions, and everything else about Claude Science, see the official Claude Science documentation , in particular:

How Claude Science uses a Slurm cluster¶

When you add an SSH host in Claude Science, it uses your existing ~/.ssh/config and authenticates with your SSH key or ssh-agent. It does not install anything on the cluster. In brief:

Probe. When you add a host, Claude Science runs a short probe that records the CPUs, memory, GPUs, environment modules, Apptainer, scratch directories, and whether sbatch exists. On a Slurm cluster it also reads the partitions. The results are saved as editable notes on the host's detail page.
Slurm submission. On a host that has sbatch, such as the Lawrencium login node, Claude Science submits jobs to Slurm by default. It does not run them on the login node itself.
Approval. Before each remote job, Claude Science shows a Run this job on <host>? card with the command and job script. Nothing is submitted until you approve it.
Job directory. On approval, the job script and inputs are copied to a job directory under the host's Scratch directory. Jobs keep running if your laptop disconnects.
Results. When the job finishes, its outputs are copied back into your Claude Science session. Files larger than about 100 MB stay on the cluster, and Claude records their paths.

Remote jobs run as you

Remote jobs run outside of Claude Science's sandbox, as your Lawrencium user, with access to everything your account can read and write on the cluster. Read the job script on the approval card before you approve it, and be careful about approving jobs for a whole project or globally.

Prerequisites¶

Claude Science installed on your computer and a Claude account on a plan that includes it. See Requirements . Claude Science needs a claude.ai account; it does not run through CBorg .
A Lawrencium user account with MFA set up, and access to at least one Slurm account (a PI Computing Allowance, condo, or recharge account). See User Accounts.
SSH access without a password prompt. Claude Science authenticates with an SSH key or ssh-agent and cannot type your PIN+OTP for you. On Lawrencium, use an SSH certificate as described below.

Step 1: Set up passwordless SSH¶

Lawrencium login normally requires your PIN+OTP. To let Claude Science connect without a prompt:

Request an SSH certificate as described in SSH Certificate.
Add a lrc-login host entry that uses the certificate to ~/.ssh/config, as shown in VS Code Remote - SSH.
Check that ssh lrc-login hostname works without asking for a password.

Certificates expire after 12 hours

When the certificate expires, Claude Science can no longer reach the cluster to submit jobs, check on them, or copy results back. Jobs already running in Slurm keep running. Request a new certificate to reconnect.

Step 2: Add the SSH host in Claude Science¶

In Claude Science, go to Settings > Compute > SSH hosts > Add SSH host.
Choose or type the alias lrc-login. The address, user, and identity file are read from your ~/.ssh/config.
In the notes field, add the Lawrencium details that Claude should know before its first job. See Host notes for Lawrencium below for a template.
Click Add. Claude Science probes the login node and saves what it finds as notes on the host's detail page. You can re-run this later with Probe.

Then, on the host's detail page, set:

Scratch directory: a directory on a shared filesystem that the compute nodes can also see, for example /global/scratch/users/your-username/claude-science. Do not use /tmp or another directory that is local to the login node. Keep in mind that your home directory has a 30 GB quota.
Concurrent job limit: Lawrencium allows at most 100 jobs per user, pending and running combined. Claude Science's default is 100, which counts only the jobs Claude submits. Set a lower limit (for example, 20) so that your own jobs and Claude's jobs together stay under the cluster limit.

Host notes for Lawrencium¶

Claude reads the notes on the host's detail page (the Details document) before it runs jobs, and it updates them as it works with the host. Accurate notes are the best way to get correct job scripts on the first try. Lawrencium jobs need a matching account, partition, and QoS. To list the combinations available to you, run the following on the login node:

sacctmgr show association -p user=$USER

Review the details document and make changes to any obvious errors. After that you can add additional information on the "User provided Context" section. A sample is below:

## User-provided context 
- Always set an account, partition and QoS. 
- Prefer es0 and lr6 for small test jobs. Use es2 for AI/ML training jobs
- Home: /global/home/users/your-username (30 GB quota)
- Never run a recursive traversal on /, /global, /global/home, 
  /global/software or /global/scratch. 
  Stay inside my own directories and bound any traversal.
- Lawrencium documentation: https://scienceit-docs.lbl.gov/llms-full.txt
- Use environment modules: `module av`, `module spider <name>`.
- Python: `module load miniforge3/25.9.1`.
- Home: /global/home/users/my-username (30 GB quota)
- Scratch: /global/scratch/users/my-username (not backed up; for temporary data)
- Never run a recursive traversal on /, /global, /global/home, /global/software
  or /global/scratch. Stay inside my own directories and bound any traversal.

For the partitions, QoS values, and GPU types on each cluster, see CPU Cluster, GPU Cluster, and Slurm Overview.

Adding the Lawrencium skill to Claude Science¶

The Lawrencium skill lets Claude answer questions about Lawrencium and Einsteinium from this documentation, for example to look up partitions, QoS values, or software modules while writing a job script. To add it to Claude Science:

Go to Settings > Skills > Add skill > Import from GitHub.
Enter the repository lbnl-science-it/scienceit-docs (or https://github.com/lbnl-science-it/scienceit-docs) and import the lawrencium-hpc skill.

Claude loads the skill automatically when a question is about Lawrencium. Imported skills don't update on their own. To get the latest version, open the three-dot menu next to the repository under Settings > Skills > Imported and select Check for updates…. See Skills in the Claude Science documentation for details.

Running jobs¶

Once the host is set up, ask Claude to run work on Lawrencium in plain language, for example:

"Run this parameter sweep on Lawrencium as a Slurm array job on lr7."
"Train this model on one H100 on es2 and plot the loss curve."

Claude writes the job script, and you approve it on the Run this job on lrc-login? card. You can approve Once, for This conversation, for This project, or Global.

Keep in mind:

Job timeout. Claude Science's default job timeout is 30 minutes. For longer jobs, tell Claude the expected run time before it submits the job.
Usage. Jobs that Claude submits are charged to the Slurm account in the job script, just like your own jobs. See Charges and Billing.
Monitoring. Claude's jobs appear in squeue -u $USER like any other job, and you can cancel them with scancel. See Monitor Jobs.
Large outputs. Outputs larger than about 100 MB stay in the job directory on scratch. Use the lrc-xfer data transfer node or Globus to move large data.

Feedback¶

Questions or problems using Claude Science with Lawrencium? Email scienceithelp@lbl.gov. For problems with Claude Science itself, see Troubleshoot Claude Science .