NVIDIA Agent Skills

The NVIDIA/skills repository is NVIDIA's official, public catalog of agent skills for using NVIDIA products from within AI coding agents. Per the provider, these are "portable instruction sets" that equip an agent with the domain knowledge needed to use NVIDIA software correctly — spanning Physical AI and robotics workflows, simulation, CUDA-X libraries, RAG pipelines, and ready-made AI blueprints. Instead of re-explaining, for every single request, how to set up a cuOpt optimization problem or generate a DeepStream pipeline, each skill captures that knowledge once in a SKILL.md file that the agent automatically pulls in once it recognizes the matching task.

Scope and origin of the skills

Per the provider, the catalog contains more than 100 skills, maintained in their respective NVIDIA product repositories and synchronized into this central repository automatically on a daily basis. Covered product families include AIQ, BioNeMo, CUDA-Q, cuDF, cuOpt, DALI, DeepStream, DOCA (with over 50 dedicated skills), Dynamo, Earth2Studio, the Holoscan SDK plus HoloHub, Jetson (BSP and device skills), medical AI skills for DICOM, segmentation, and image synthesis, the NeMo suite (AutoModel, Fabric, Relay, Retriever, RL), the TAO Toolkit with over 100 vision models, as well as Video Search and Summarization and Warp. Example individual skills include cuopt-routing-api-python, deepstream-generate-pipeline, nemo-retriever, and tao-train-image-classification. Because the skills originate directly from the product teams rather than being reverse-engineered by third parties, the catalog reflects the actual capabilities of the underlying NVIDIA software fairly closely.

Signature mechanism and governance

One thing that sets this apart from many other skill collections is its cryptographic hardening: per the provider, every published skill carries a detached signature file (skill.oms.sig) generated via the ONNX Model Signing (OMS) format, which can be verified against an NVIDIA trust anchor (nv-agent-root-cert.pem). According to the README, the automated sync pipeline additionally enforces signature-drift detection and missing-artifact checks. Beyond the SKILL.md itself, the publishing process also requires a skill-card.md with governance metadata and mandatory Tier-3 evaluation datasets, optionally supplemented by a BENCHMARK.md with performance data. This positions the repository explicitly as a reference implementation for "verified" agent skills rather than a collection of freely copyable snippets.

Supported coding agents and installation

Installation runs through the vendor-neutral "skills" CLI and, per the provider, supports several agents at once, including Claude Code, Codex, Snowflake Cortex (CoCo), Cursor, and the Kiro CLI. The base command is:

npx skills add nvidia/skills

Individual skills can be installed without prompts, for example npx skills add nvidia/skills --skill cuopt-numerical-optimization-api --yes, and targeted at multiple agents at once via repeated --agent flags, e.g. --agent claude-code --agent codex. Running npx skills add nvidia/skills --list shows the full, searchable skill catalog, while npx skills update keeps already-installed skills current.

License and cost

Per the license files, the repository is dual-licensed: source code under Apache-2.0 (LICENSE-APACHE), documentation and the skill content itself under CC-BY-4.0 (LICENSE-CC-BY-4.0). Use is entirely free of charge, with no paid tier for the catalog itself — any cost can only arise indirectly from using the underlying NVIDIA cloud services or hardware referenced by an individual skill, which is noted within that skill itself.

Who benefits from this repository

The catalog is especially valuable for teams already working with NVIDIA products such as CUDA-X, Isaac/Jetson robotics, Omniverse-adjacent simulation, or the NeMo suite who want to equip their AI coding agents with vendor-verified rather than guessed domain knowledge. It's also of interest to organizations with governance requirements, since every skill can be traced back to its official NVIDIA source via the OMS signature instead of relying on an unverified community write-up. Anyone not using NVIDIA-specific software, however, will find more generic collections such as the Anthropic Agent Skills or the Upstash Agent Skills a better fit.