NERSC AI coding agent tools

This repository provides mirrored user-level artifacts for running AI coding agents safely and effectively on NERSC Perlmutter. The claude/ and codex/ directories contain equivalent artifacts adapted to Claude Code and Codex, respectively.

The package currently contains:

• Perlmutter agent guidance: a general CLAUDE.md or AGENTS.md template that describes shared login-node best practices, Slurm approval and requirements, NERSC software and filesystem conventions, and user-specific extension points.

In the future, we also plan to include:

• perlmutter-compute skill: workload-routing guidance and helper scripts that decide whether work belongs on a login node, in a new (or active) interactive job, in a batch job, or in a workflow environment. The skill also supports login-node preflight and monitoring, allocation checks, and preparation of reviewed Slurm commands and job scripts.
• Additional capability-specific skills to help agents act more effectively on NERSC systems. Examples may include containers, GPU programming, IRI API, PyTorch training, etc. These skills should be added based on your specific needs.

Clone this repo in your $HOME directory to install the artifacts and pull future updates. For detailed installation and update instructions, see:

• Claude Code installation
• Codex installation

These artifacts are per-user coding agent customizations. They are not system-wide NERSC policy. We recommend that all users working with AI coding agents use our CLAUDE.md or AGENTS.md and perlmutter-compute skill (when it is available), as they help agents run safely and effectively on NERSC systems.

In addition to these user artifacts, we currently have the following managed hooks deployed at NERSC for administrative-level safeguards:

• Filesystem guard: prevents Claude Code and Codex from performing broad filesystem traversals.

For more information on AI coding agent best practices at NERSC, see our documentation:

• AI coding tools at NERSC