create-pod

name	create-pod
description	Create RunPod GPU/CPU pods from runpod.toml. "create pod", "launch pod", "pod 立てて" などで起動。
metadata	{"author":"pokutuna"}
allowed-tools	["Bash(uv run --script ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/scripts/create_pod.py:*)","Bash(uv run --script ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/scripts/create_cpu_pod.py:*)"]

RunPod Pod Creation

Create RunPod pod instances from a runpod.toml configuration file.

Important

• Pods incur costs while running. If creation or SSH fails, confirm with the user and stop/remove the pod
• CPU pods have a max container disk of 20GB
• CPU pods: if datacenter_id in runpod.toml is not accepted by the REST API, use --datacenter "" to auto-place near the Network Volume
• CPU pod defaults are tiny: cpu3c + --vcpu 2 = 4 GB RAM. Large file downloads (e.g. hf download of multi-GB model shards) will SIGKILL (exit 137) due to cgroup memory limit. For HF model downloads, use --cpu-flavor cpu3g --vcpu 4 (16 GB) or larger.
• Use the non-deprecated runpodctl commands (legacy forms still work but emit a deprecation warning):
  • Create pod: runpodctl pod create (NOT runpodctl create pod)
  • List pods: runpodctl pod list (NOT runpodctl get pod)
  • SSH info: runpodctl ssh info <pod-id> (NOT runpodctl ssh connect <pod-id>)
  • GPU types: runpodctl gpu list (NOT runpodctl get gpus)
• create_pod.py requires runpodctl >= 2.1.7 (for --network-volume-id on the new pod create form)
• Do NOT wrap create_pod.py in a shell while retry loop. Use the built-in --retry flag — wrapping creates duplicate pods if your success/failure detection is wrong. The script also refuses to create when a pod with the same name is already running (override with --allow-duplicate).

Prerequisites

• runpodctl CLI installed and configured (~/.runpod/config.toml)
• runpod.toml in the working directory (or specify with -c)

Setup

If runpod.toml does not exist yet:

cp ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/templates/runpod.toml ./runpod.toml
mkdir -p scripts/runpod
cp ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/templates/init.sh ./scripts/runpod/init.sh

Edit runpod.toml with your settings. Register secrets at https://console.runpod.io/user/secrets and reference them as {{ RUNPOD_SECRET_XXX }} in [env].

Choosing the Script

Request	Script	Reason
GPU pod (default)	create_pod.py	Wraps runpodctl CLI
CPU-only pod	create_cpu_pod.py	Uses REST API (runpodctl does not support CPU pods)

Both scripts share the same runpod.toml. CPU script ignores GPU-specific fields (gpu_type, gpu_count).

GPU Pod

uv run --script ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/scripts/create_pod.py                       # Create a GPU pod
uv run --script ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/scripts/create_pod.py --ssh                 # Create and SSH connect
uv run --script ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/scripts/create_pod.py --ssh --retry         # Wait for stock, then SSH in
uv run --script ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/scripts/create_pod.py --dry-run             # Show command only
uv run --script ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/scripts/create_pod.py --gpu "RTX 5090"      # Override GPU type

For low-stock GPUs, prefer --ssh --retry over wrapping the script in a shell loop. See --help for --retry-interval / --retry-max.

CPU Pod

uv run --script ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/scripts/create_cpu_pod.py              # Create a CPU pod
uv run --script ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/scripts/create_cpu_pod.py --ssh        # Create and SSH connect
uv run --script ${CLAUDE_PLUGIN_ROOT}/skills/create-pod/scripts/create_cpu_pod.py --dry-run    # Show request body only

See --help for all options including --cpu-flavor, --vcpu, --datacenter.