dsh-remote-gpu-monitoring

Managing a fleet of GPU servers typically means ssh-ing into each box one by one, running nvidia-smi by hand, and juggling multiple terminal panes. dsh-remote-gpu-monitoring collapses that entire workflow into a single live panel in the DeepSeek Harness sidebar: on startup it auto-discovers every concrete alias in your ~/.ssh/config, issues one fixed read-only query per host, and renders per-card model, memory, utilization, temperature, and power draw with threshold-based color coding.

The panel lets you toggle individual servers on or off; unchecked hosts stop being polled immediately. All browser sessions share a single host-side cache backed by ControlMaster socket reuse and exponential backoff, so a five-second tick is a trivial exec over an existing connection. On the model side, two agent tools — gpu_overview (zero-SSH cached snapshot) and gpu_refresh (immediate collection pass) — turn questions like "which server has a free card?" into a single tool call instead of seven SSH round-trips.

Built for developers and research teams who run multi-node GPU stacks, or anyone who wants a real-time cluster status board right inside their conversation interface. No extra services, no config files — install and go.