runpod__runpod-plugins-official/plugins/runpod

Runpod

One entrypoint for AI agents to manage GPU workloads on Runpod — pods, serverless endpoints, jobs, templates, and volumes — via the Runpod MCP server, runpodctl, and flash, with conceptual guidance and worked golden paths.

This plugin bundles six skills, the hosted Runpod MCP server config (.mcp.json), and reference material.

Which skill?

Start with runpod — the router. It reads your task and points to the right lane below. If you already know the lane, go straight to it.

Skill	Use it for
runpod	Router / entrypoint. Start here when the right skill is unclear.
runpod-mcp	Manage infra (pods, endpoints, jobs, templates, volumes, catalog, billing) via the Runpod MCP server's structured tool calls.
runpodctl	Manage infra from the CLI, plus Hub deploys, file transfer (send/receive), SSH, and doctor setup.
flash	Write and deploy your own code on Runpod serverless — @remote/@Endpoint, flash dev, flash deploy.
companion-clis	Prerequisite CLIs: hf (models), docker (images), gh (repos/releases), aws (S3 to volumes).
runpod-usage	Concepts — how pods/serverless work, building containers, storage, GPU selection, gotchas.

runpod-mcp vs runpodctl: both drive the same Runpod API for the same infra CRUD. Prefer runpod-mcp when its tools are connected in your session; use runpodctl for the terminal, Hub, file transfer, SSH, or doctor.

The development loop

Any "get X running on Runpod" task follows one loop (in runpod-usage): **decide pod vs serverless → prefer a prebuilt template/Hub worker over from-scratch → provision → verify with a real request ("Running" ≠ "ready") → deliver → cost-guard

• teardown.** See skills/runpod-usage/reference/development-loop.md.

The skills/runpod/golden-paths/ folder holds worked, end-to-end reference tasks (Ollama, ComfyUI, Whisper, …) — acceptance scenarios, not installed skills (they have no SKILL.md, so agents don't load them). They live under the runpod router skill that indexes them, so they travel with it on a single-skill install.

Setup

Everything unifies on a single RUNPOD_API_KEY (https://runpod.io/console/user/settings):

runpodctl doctor          # CLI: store the key + SSH

The hosted MCP server (bundled in .mcp.json) is the exception — it uses the "Sign in with Runpod" OAuth flow, so no key is stored on disk (see runpod-mcp). Companion CLIs (hf, gh, docker, aws) use their own credentials.

Usage

Ask your AI agent:

• "Create a pod with an RTX 4090"
• "Deploy a serverless endpoint from this image"
• "Which GPU should I use for a 13B model?"
• "Write an @remote function and run it on a GPU"
• "Download a model, containerize it, and deploy it"

URLs

• Pod: https://<pod-id>-<port>.proxy.runpod.net (e.g. https://abc123xyz-8888.proxy.runpod.net)
• Serverless: https://api.runpod.ai/v2/<endpoint-id>/{run|runsync|health|status/<job-id>}

More in skills/runpod-usage/reference/networking.md.

Structure

skills/
  runpod/            router / entrypoint
    golden-paths/    worked end-to-end reference tasks (indexed by the router)
  runpod-mcp/        Runpod MCP server (structured tool calls)
  runpodctl/         Runpod CLI (+ Hub, transfer, SSH, doctor)
  flash/             write & deploy your own code (@remote)
  companion-clis/    hf / gh / docker / aws prerequisites
  runpod-usage/      concepts + reference/*.md
.mcp.json            hosted Runpod MCP server config

License

Apache-2.0