NVIDIA/skills

What It Is

NVIDIA/skills is a catalog repository: each entry is a portable Agent Skill — an instruction set rooted at a SKILL.md file — that teaches an AI coding agent how to use an NVIDIA product. Covered areas include Physical AI and robotics, simulation, CUDA-X libraries, RAG and AI Blueprints, and platform tools such as Holoscan, DeepStream, DOCA, cuOpt, cuDF and BioNeMo. It is aimed at developers who already drive Claude Code, Codex, Cursor, Kiro, Snowflake CoCo or similar agents and want those agents to produce correct NVIDIA API usage instead of generic guesses.

The catalog is not the skills' home. Skills are authored in their respective product repositories and mirrored here daily by an automated sync pipeline. A skill reaches skills/ only after the pipeline confirms it carries a detached OMS signature (skill.oms.sig), a skill-card.md governance card and a Tier-3 evaluation dataset; when evaluation runs produce a BENCHMARK.md, that ships alongside the skill.

Quickstart

Prerequisites: Node.js with npx available, and a skills CLI at v1.5.16 or newer. Install the Skill Finder first, which can then locate and install other skills for your use case:

npx skills add nvidia/skills --skill nvidia-skill-finder --yes

When you already know the skill name, install it directly. For example, to add cuOpt's numerical optimization API skill:

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

To target a specific agent, add --agent; the flag can be repeated to install into several agents at once:

npx skills add nvidia/skills \
  --skill cuopt-numerical-optimization-api \
  --agent claude-code \
  --agent codex

Supported --agent values in the docs include claude-code, codex, cortex (Snowflake CoCo), cursor and kiro-cli. After installing, the skill becomes available the next time your agent loads skills and meets a relevant task. In Claude Code, run /reload-skills to load it into the current session without restarting.

Useful companion commands: npx skills add nvidia/skills --list to browse before installing, npx skills update to refresh installed skills, npx skills list to see what is installed and npx skills check to preview what is out of date. Running update interactively also flags skills removed or merged upstream and offers to delete stale local copies. Non-interactive, global and agent-specific installation plus manual copying fallbacks are covered in the advanced installation guide.

Verifying Skills

Each published skill ships with a detached OMS signature that can be checked against NVIDIA's trust anchor, nv-agent-root-cert.pem, using the model-signing package:

pip install model-signing
model_signing verify certificate SKILL_DIR \
  --signature SKILL_DIR/skill.oms.sig \
  --certificate_chain nv-agent-root-cert.pem \
  --ignore_unsigned_files

A successful verification confirms the skill contents have not been modified since NVIDIA signed them. Signature layout, trust pipeline and policy options are described in the signing documentation.

Constraints and Compatibility

The CLI version requirement is not cosmetic: on v1.5.15 and earlier, skills may install but fail to appear in Claude Code, so npx skills@latest add ... is the recommended form. Skills follow the Agent Skills specification — SKILL.md at the root, YAML frontmatter with required name and description, and progressive disclosure so only lightweight metadata loads at startup. Compatibility of --agent targets and validation of authored skills are handled by the skills CLI and its reference library; the CLI's own supported-agent table is the authority for clients beyond those named here.

Licensing is split: source code under Apache 2.0, documentation and skills under CC BY 4.0. Some skills retrieve external materials (models, datasets, services) that are not distributed with the repository and are governed by their own terms; the README places responsibility for reviewing and complying with those terms on the user.

Catalog maintenance follows a clear split: content bugs in a specific skill go to that product's source repository (the per-product issue table in the README maps each one), while catalog, sync, signing and documentation problems go to this repository's issue templates. Questions belong in Discussions, and security vulnerabilities follow SECURITY.md rather than a public issue.

HotRepo's take

On the Python weekly list dated 2026-11-14, the repository recorded 92 new stars and a best position of 17, reaching 3,988 stars and 524 forks in total. One weekly snapshot alone does not establish a trend, but the placement is consistent with the catalog's structure: the value is the curated, signed and benchmarked breadth across NVIDIA product lines rather than any single skill. The governance design is the more interesting signal — signatures, skill cards, evaluation datasets and sync-time artifact gates are exactly the controls a general-purpose agent marketplace lacks, and they matter because an installed skill is executable guidance that shapes what an agent writes. No GitHub daily rank is recorded for this repository, so the weekly figure should not be read as an all-language daily placement.