skill · curated · nvidia · cuda · physical-ai · developer-tools · claude-code · Aug 24, 2026 · 2 min read
nvidia/skills: official agent skills for CUDA-X, Physical AI, and robotics
What it is
nvidia/skills is NVIDIA's official catalog of Agent Skills — portable instruction sets that teach your coding agent how to use NVIDIA software correctly. One command gives Claude Code, Codex, or Cursor working knowledge of 50+ products: CUDA-X libraries, Physical AI workflows, robotics simulation via Warp, RAG pipelines, medical imaging (MONAI), quantum computing (CUDA-Q), and video analytics (DeepStream, Holoscan). Every skill in the catalog is OMS-signed and auto-synced daily from upstream NVIDIA product repos.
When it earns its keep
The value proposition is narrow but sharp: if you're already in the NVIDIA ecosystem and your agent keeps guessing at the right initialization sequence, the right cuOpt API surface, or the right way to wire a DeepStream pipeline — this is the fix. The skills encode workflow knowledge that isn't obvious from reading the docs linearly: correct call order, the footguns, the flags that only matter on certain driver versions.
The breadth is wider than you'd expect for a product-specific catalog. Warp (differentiable simulation for robotics), CUDA-Q (quantum computing), MONAI (medical imaging), cuDF (GPU DataFrames) — they're all here, updated daily from the same repos that ship the actual SDKs. That's a meaningful difference from a community skill that drifts stale.
The honest caveat
These skills teach your agent the language — they don't provide the interpreter. If you don't have CUDA drivers installed, licensed SDKs, or actual NVIDIA GPUs in the picture, adding this catalog is just vocabulary for a conversation that can't happen. The skills are also silent on everything outside the NVIDIA stack, so they won't help you decide *whether* to reach for cuOpt vs. a vanilla solver — just *how* to use cuOpt once you've decided.
One practical note: the catalog requires skills CLI v1.5.16 or newer. On older CLIs the install might silently succeed but never surface in your agent. Run npx skills@latest add nvidia/skills to be safe — plain npx skills on an older install can fool you.
Install
npx skills@latest add nvidia/skillsThe catalog is large enough that bulk-installing everything adds noise for projects that only touch one or two NVIDIA subsystems — the interactive browser that skills add launches lets you pick only what you need.
Credit
Built and maintained by NVIDIA. Open on GitHub at github.com/nvidia/skills — skills are OMS-signed, versioned, and carry governance metadata (provenance, quality tier, behavioral boundaries) alongside the actual instructions. 3.1k stars as of this writing.