NVIDIA Agent Skills

NVIDIA Agent Skills is an official catalog of portable, NVIDIA-verified skills that teach AI coding agents how to use NVIDIA software. Skills are maintained in their respective product repositories and mirrored into this repository daily by an automated sync pipeline. They can be installed into Claude Code, Codex, Cursor, Kiro and other agents compatible with the open Agent Skills specification.

What It Is

Agent skills are portable instruction sets, each a directory with a SKILL.md file, that an agent loads when a matching task comes up. The catalog covers CUDA-X libraries such as cuDF, cuOpt and CUDA-Q, along with RAG and AI Blueprints, NeMo training and inference tooling, Dynamo inference serving, DeepStream and video search, Jetson, DOCA, TAO Toolkit, and Omniverse and Physical AI workflows. For example, asking an agent to solve a linear programming problem with cuOpt guides it through the cuOpt Python API.

Installation Workflow

Skills are installed with the Vercel-maintained skills CLI through npx skills add nvidia/skills. The CLI prompts for a skill and a destination, or accepts a skill name and --agent flags to target specific agents in one command. Further commands list the catalog, update installed skills, and flag skills that were removed or merged upstream. The docs say the NVIDIA skills plugin will also be discoverable through the Claude Code and Codex marketplaces, with Hermes Skillhub and ClawHub planned.

Trust Controls and Verification

The docs describe verified skills as release artifacts. Each is cataloged from its owning product team, scanned for software and agent-native risks such as prompt injection and hidden instructions, and evaluated against a task set with and without the skill. Each skill is signed with a detached skill.oms.sig and documented with a skill card. Consumers can verify signatures against the published NVIDIA root certificate using the model-signing tool. A BENCHMARK.md report ships with skills when evaluation runs produce one.

Licensing

The repository is dual-licensed: source code under Apache-2.0 and documentation and skills under CC-BY-4.0.