nvidia/skills: NVIDIA's Official Verified Skill Catalog for Coding Agents

Your coding agent has probably seen plenty of pandas, but ask it to formulate a vehicle-routing problem for cuOpt or flash a Jetson BSP image and it starts hallucinating flags. nvidia/skills is NVIDIA's official answer: a catalog of NVIDIA-verified Agent Skills that teach AI agents how to use NVIDIA software — physical AI and robotics workflows, simulation, CUDA-X libraries, RAG blueprints, and platform tools. The repo sits at 3.1k stars under an Apache-2.0 + CC-BY-4.0 license.

Why This Skill Matters

Two design choices set this catalog apart. First, it is a mirror, not a monolith: each skill is maintained in its respective product repo and synced here daily via an automated pipeline, so the cuOpt guidance tracks the cuOpt team, not a fork that rots. Second, the verification framing — NVIDIA publishes these as verified skills with a capability-governance blog post behind them, and the catalog even ships a skill-card-generator skill that produces a governance card for a skill directory.

The catalog spans dozens of product areas: GPU-accelerated optimization (cuOpt for vehicle routing, linear and quadratic programming), data science (cuDF, cuPyNumeric), quantum computing onboarding (CUDA-Q), robotics and Jetson board support, large-scale training (Megatron-Core, the NeMo family), medical imaging on MONAI, RAG deployment blueprints, and more.

Installation

The default skills CLI flow works everywhere:

npx skills add nvidia/skills

The CLI prompts you to choose a skill and an install destination — no cloning or hand-copying folders. One requirement from the README: a current skills CLI, v1.5.16 or newer. Installing via npx skills@latest add nvidia/skills always gets the latest; older CLIs may install skills that then fail to appear in Claude Code.

To install one skill without prompts, name it:

npx skills add nvidia/skills --skill cuopt-numerical-optimization-api --yes

And to target a specific agent, use --agent (supported targets include claude-code, codex, cursor, kiro-cli, and cortex for Snowflake CoCo) — repeat the flag to install into several at once.

Real Workflow: Solve an Optimization Problem With cuOpt

1. Install the cuOpt skills:

npx skills add nvidia/skills --skill cuopt-routing-api-python --yes

2. In Claude Code, run /reload-skills to load newly installed skills in the current session.
3. Prompt the agent:

Solve a linear programming problem with cuOpt

The skill guides the agent through the cuOpt Python API instead of inventing endpoint names. The same pattern works for the formulation and server API skills if you need QP modeling or remote deployment.

Real Workflow: Browse the Catalog Before Installing

Not sure what exists yet? List without installing anything:

npx skills add nvidia/skills --list

Since new skills land continuously and existing ones get renamed or consolidated, keep what you have installed fresh with npx skills update — interactively it also flags removed or merged skills and offers to clean up stale copies. npx skills check previews what is out of date first.

Tips

• Install selectively per project — a robotics project needs Jetson and Holoscan skills, not the NeMo training set.
• Full documentation lives at docs.nvidia.com/skills, and each product row in the README links to its upstream repo's issues and contributing guides.
• Skills load the next time your agent encounters a relevant task; you do not need to invoke them by hand.
• For anything non-interactive — global installs, removals, manual copying — the README links an advanced installation guide.

When Not to Use This

Every skill here is about NVIDIA software. If your stack is vanilla Python, Node, or cloud services without GPU acceleration, there is nothing in this catalog for your daily work — and installing it wholesale just adds context your agent will never trigger.

See the leaderboard for more skills.