runpod-templates

Official Runpod templates

Prefer an official template over building an image. Runpod maintains prebuilt images for common workloads; deploying one is a create + poll, with no SSH install step. This skill is the reference for what those images actually are. It does not manage infrastructure — deploy with runpod-mcp (if connected) or runpodctl.

Terms — "template" vs the Hub

Two different products share the word "Hub", and conflating them derails discovery:

Same word, disjoint catalogs: hub search comfyui returns only the serverless worker and none of the official ComfyUI pod templates, and template search returns no Hub repos. In the REST v2 catalog, pod vs serverless templates are told apart by each entry's serverless flag.

What "official" means, and how to find them

A template is Runpod-maintained when it reports isRunpod: true. Anything else in search results is community-published: usable, but not covered by this skill and not version-guaranteed.

runpodctl template list --type official   # the whole official set (14 as of 2026-08-25)
runpodctl template search <name>          # by name; check isRunpod: true
runpodctl template get <template-id>      # image, ports, portsConfig, disk + full readme

template get returns the template's readme — the description maintained alongside the image, and the authoritative answer when it disagrees with this skill. Take image tags, ports, and env from these commands, not from these files: the templates ship on their own release train and move faster than this repo. The reference files record the shape and the gotchas, which is the expensive part to rediscover.

⚠️ runpodctl hub ... is not how you find pod templates — see Terms.

To hand a user a Console link, use https://console.runpod.io/hub/template/<template-id>.

REST v2 has no template search — only slices you filter yourself

REST v2 has no query/search parameter for templates, so every "search" is a client-side filter over a fetched slice:

official (14) and verified (3) are fully enumerable; community is capped at 100 with pagination explicitly unsupported. A community template past that cap never enters a slice the API can return, and runpodctl template search reads those same server-side slices — so it has no view past the cap either. If a user names one you cannot find, ask them for the template id rather than concluding it does not exist.

isRunpod is a GraphQL/v1 field with no v2 equivalent — in v2, "official" is the source slice you requested, not a flag on the record.

Catalog entries carry allowedCudaVersions, which is the reliable way to pick between CUDA-line variants of the same template — pair it with runpodctl pod create --min-cuda-version instead of inferring from the GPU name.

The templates

Add a template by adding one file here and one row above. Do not create a new skill per template — the task shape is identical and every registered skill costs context in every session.

How to use a template reference

Each file under reference/ answers the same questions in the same order, so an agent can skim to the one it needs:

1. Identity — template name, id, image, upstream source repo.
2. Variants — CUDA/arch/GPU-generation splits and which to pick.
3. Ports and credentials — what is exposed and any default login.
4. Autostart — what is already running on boot, with the exact command.
5. Paths — where the app, its config, and its data live (and what a network volume mount changes).
6. Readiness — how long first boot takes and the exact signal for "actually serving". Running is never the signal.
7. What does not ship — the gap between "booted" and "usable", and how to close it programmatically rather than by clicking.
8. Sizing — container disk and minimum VRAM.
9. Version pinning — how to hold a known-good version.

One file is the exception: reference/comfyui-model-repair.md is a usage guide — a ComfyUI workflow repair procedure that drives the scripts in scripts/ — not a 9-question template reference.

Routing onward — this skill is the hub, not the destination

A user asking about a template rarely wants the reference file itself; they want a template deployed, fixed, or replaced. Land here, identify which, and hand off:

Reference files here answer what is in the image; they never duplicate a walkthrough or a repair procedure — they point at the owner.