Skip to content

On the Cluster (Login Nodes)

In this mode you SSH into a cluster login node and run a command-line harness right there. The four CLI harnesses — Claude Code, Codex, Gemini CLI, and opencode — install and run headlessly on Linux, so they work over SSH. Warp does not run here; it is a desktop application (see Warp below).

The login-node rule applies to your agent

A login (front-end) node is a shared, multi-tenant machine meant for editing, compiling, submitting jobs, and light pre- and post-processing — not for running computation. Launching a harness there is fine: the agent process itself is lightweight (mostly network calls). But anything heavy the agent then wants to do must go through Slurm, exactly as if you were doing it by hand.

Important

Do NOT run large, long, multi-threaded, parallel, or CPU-intensive jobs on a front-end login host. All users share the front-end hosts, and running anything but the smallest test job will negatively impact everyone's ability to use the cluster. Always use SLURM to submit your work as a job.

In practice: instruct your agent to compile modestly, test on tiny inputs, and submit real runs with sbatch or sinteractive — with a correct account (-A, discovered via slist), partition (-p), QOS (-q), and an explicit --time limit. The shared context files RCAC publishes already tell an agent this, and the per-harness settings deny the most dangerous operations, but the responsibility is ultimately yours.

Where agents may write

Point an agent's working files at your scratch space, not your home directory. Scratch is the large, high-performance filesystem intended for job I/O; find it with findscratch or the $RCAC_SCRATCH environment variable. Keep in mind scratch is not backed up and is purged after a period of inactivity (the window varies by cluster), so move anything you want to keep to durable storage.

Sandboxes are weak on shared nodes — don't rely on them

The harnesses ship OS-level sandboxes (Codex uses bubblewrap, Gemini uses Docker/Podman, Claude Code uses Linux namespaces). These are frequently unavailable or disabled on shared login nodes, so do not count on them. Containers are not a safety net either: RCAC uses Apptainer (not Docker), and its configuration auto bind-mounts /home, /depot, and /scratch into the container — those mounts stay writable, so an agent in a container can still edit your real files. Treat the harness's permission/approval rules as your primary control: keep destructive operations (rm -rf, sudo) denied and set writable roots to $RCAC_SCRATCH.

Install and run each harness

Each CLI installs into your user space (home directory) — no elevated privileges needed. Run these on a cluster login node after you SSH in.

1
2
3
4
5
# native installer (or: npm install -g @anthropic-ai/claude-code)
curl -fsSL https://claude.ai/install.sh | bash

claude                 # interactive
claude -p "…"          # headless / non-interactive

1
2
3
4
5
# shell installer (or: npm install -g @openai/codex)
curl -fsSL https://chatgpt.com/codex/install.sh | sh

codex                  # interactive
codex exec "…"         # headless / non-interactive

1
2
3
4
npm install -g @google/gemini-cli

gemini                 # interactive
gemini -p "…"          # headless (avoid --yolo on shared nodes)

1
2
3
4
5
# install script (or: npm install -g opencode-ai)
curl -fsSL https://opencode.ai/install | bash

opencode               # interactive TUI
opencode run "…"       # non-interactive

To connect these harnesses to the RCAC MCP servers and wire in the shared context and permission policy, see Shared Context & Settings.

Warp

Warp is RCAC's recommended harness for most users — but it is a desktop application, not a headless CLI, and cannot be installed on a login node. There is no login-node path for Warp, and you should not try to fabricate one.

Instead, run Warp on your own workstation and SSH into the cluster from there; its Agent Mode then rides your live SSH session — reading output and issuing commands in the session you opened — while the app stays local. That is the recommended setup for Warp; see Local, Targeting the Cluster.

Back to Running Agents.