Running Agents¶
There are two ways to run an agentic coding tool against an RCAC cluster, and this section is organized around that choice. Both are legitimate; which one fits depends on your harness and how you like to work.
- On the cluster (login nodes). You SSH into a cluster login node and run a command-line harness (Claude Code, Codex, Gemini CLI, or opencode) there. The agent lives on the front end alongside your shell, so it sees the real filesystem, modules, and scheduler directly — but it is bound by the same login-node etiquette you are.
- Locally, targeting the cluster. The harness runs in your own environment — your laptop or workstation — and reaches the cluster over your existing SSH access. Two mechanisms fit here: most harnesses (Claude Code, Codex, Gemini CLI, opencode) connect through the RCAC MCP servers, which bridge to the cluster over SSH so the agent works through tool calls; Warp is a desktop terminal in which you SSH into the cluster yourself and its Agent Mode rides your live session. Either way, nothing new is installed on the cluster and no new credentials are created.
-
On the Cluster (Login Nodes)
Run a CLI harness directly on a cluster front end. Covers install, the login-node compute rules, and where agents may write.
-
Local (over SSH)
Run the harness on your own machine and reach the cluster over SSH — no new credentials, no hosted infrastructure. The recommended setup, and the only way to use Warp.
Back to Agentic AI.