NVIDIA Agent Skills

What It Does

NVIDIA Agent Skills is an official, curated catalog of 108 agent skills published by NVIDIA Corporation. The skills teach AI agents how to use NVIDIA's CUDA-X libraries, AI Blueprints, and platform tools: GPU-accelerated optimization (cuOpt), large-scale model training (NeMo, Megatron-Core), inference serving (Dynamo), robotics and simulation (Omniverse, PhysX), quantum computing (CUDA-Q), and more. Skills are maintained in their respective product repositories and mirrored to a central catalog daily via an automated sync pipeline. The repo follows the Agent Skills specification (agentskills.io) and supports installation via the standard npx skills add CLI into Claude Code, Codex, Cursor, and Kiro.

The Good

Cryptographic signing and a formal trust pipeline. Every published skill carries a detached OMS-format signature (skill.oms.sig) verifiable against NVIDIA's root certificate (nv-agent-root-cert.pem). The signing uses Sigstore-style DSSE envelopes with SHA-256 digests of every file in the skill directory. This is the first agent skill collection in the ecosystem with this level of supply-chain integrity. The trust pipeline documentation (7 pages across 3 docs) covers scanning, signing, and skill cards as three complementary layers.

3-tier evaluation framework with published benchmarks. Each skill ships with a BENCHMARK.md that records results from NVSkills-Eval across 5 dimensions: security, correctness, discoverability, effectiveness, and efficiency. The evaluation tests skills against multiple agents (Claude Code, Codex) and reports per-dimension scores with and without the skill. Of 108 skills, 61 pass and 44 fail their benchmarks. Publishing failures alongside passes is a genuine transparency signal that no other skill collection in the ecosystem matches.

Governance skill cards for every skill. All 108 skills include a skill-card.md documenting owner, license, use case, deployment geography, known risks and mitigations, output shape, evaluation agents used, and ethical considerations. The skill-card specification (documented in docs/skill-cards.mdx) provides concrete guidance on writing actionable risk statements, with before/after examples. This is the only skill collection that ships per-skill governance metadata at this level of detail.

High skill quality with strong median depth. Across 108 SKILL.md files, the median length is 204 lines. The largest skill (vss-deploy-dense-captioning) runs 500 lines. Skills include structured sections for problem-type selection (e.g., LP vs MILP vs QP decision tables), code examples, error handling guidance, and cross-references to stable documentation. The 10 skills under 30 lines are primarily index/navigation skills (nemoclaw-user-overview, nemoclaw-user-reference) rather than stubs.

Well-organized multi-product architecture. The components.d/ directory holds 27 product registration files. The plugins.d/ system builds curated plugin bundles from the full catalog, with a nvidia-skills plugin that ships a hand-picked subset covering the most important NVIDIA products. The skills.sh.json groups skills into 9 categories (Agentic AI, Training AI, Inference AI, Decision Optimization, etc.) with descriptions. The CI pipeline validates plugins on every PR, ensuring the generated tree matches the declarations.

Rich per-skill assets. Skills average 14 files each. 71 skills ship with a references/ directory containing supplementary documentation. 92 skills include evals.json evaluation datasets. 10 skills include assets/ directories with runnable code examples (Python scripts, data files). The cuopt-numerical-optimization-api-python skill includes 8 complete example projects (LP basic, MILP basic, portfolio optimization, least squares, warm start, etc.).

The Bad

44 of 108 skills fail their own benchmarks. While publishing failures is transparent, the failure rate is 41%. The NeMo-MBridge performance skills are especially hard-hit: 17 of 18 MoE/performance tuning skills fail. Several NemoClaw skills also fail (configure-inference, configure-security, deploy-remote, manage-policy, manage-sandboxes, overview, reference). These are not trivial misses. They represent skills that either did not improve agent performance or introduced regressions on evaluation tasks. Readers installing a failing skill may get worse results than no skill at all.

Domain specificity limits audience. These skills are for developers working with NVIDIA hardware and software. The cuOpt skills require an NVIDIA GPU and cuOpt installation. The NeMo/Megatron-Core skills require multi-GPU clusters. The Dynamo skills require Kubernetes deployments. The Omniverse skills require NVIDIA Omniverse. Developers using AMD, Intel, or Apple Silicon will find almost nothing applicable. This is not a criticism of quality, but it limits the install base to the NVIDIA ecosystem.

Four stub skills under 20 lines. While most skills are substantial, 4 skills are under 20 lines: nemoclaw-user-agent-skills (10 lines), nemoclaw-user-configure-security (16 lines), nemoclaw-user-overview (17 lines), nemoclaw-user-reference (18 lines). These are navigation/index skills that point to references rather than providing standalone value. They inflate the skill count without adding proportional depth.

No test suite for the catalog itself. The CI pipeline validates plugin structure and author identity (DCO check), but there is no test suite that validates SKILL.md frontmatter consistency, cross-reference resolution, or skill-card completeness across all 108 skills. The validate-plugins.yml workflow checks the generated plugin tree, but not individual skill metadata quality.

Smoke Test Results

Smoke testing was done on macOS (host) against the cloned repository and via live npx skills add install.

Structural validation (host-based, cloned repo)

$ cd /tmp/nvidia-skills-review
$ find ./skills -name "SKILL.md" | wc -l
108

$ find ./skills -name "skill-card.md" | wc -l
108

$ find ./skills -name "*.sig" | wc -l
108

$ find ./skills -name "BENCHMARK.md" | wc -l
108

$ grep -l "Overall verdict:" $(find ./skills -name "BENCHMARK.md") | xargs grep "Overall verdict:" | grep -c "PASS"
61

$ grep -l "Overall verdict:" $(find ./skills -name "BENCHMARK.md") | xargs grep "Overall verdict:" | grep -c "FAIL"
44

$ find ./skills -name "evals.json" | wc -l
92

Pass rate: 6 of 6 structural checks. All 108 skills have SKILL.md, skill-card.md, BENCHMARK.md, and signature files. 92 of 108 have evals.json datasets.

Run B. Install via npx skills add

$ cd /tmp/nvidia-skills-test && npx --cache /tmp/nvidia-skills-npx-cache \
 skills add nvidia/skills \
 --skill cuopt-numerical-optimization-api-python \
 --agent claude-code --yes

 Found 108 skills
 Selected 1 skill: cuopt-numerical-optimization-api-python
 Security Risk Assessments:
 Gen: Safe Socket: 0 alerts Snyk: Med Risk
 Installed 1 skill -> .claude/skills/cuopt-numerical-optimization-api-python

$ ls .claude/skills/cuopt-numerical-optimization-api-python/
 SKILL.md BENCHMARK.md skill-card.md skill.oms.sig
 assets/ benchmark/ evals/ references/

Pass rate: 1 of 1 install check. The npx install cloned the repo, listed all 108 skills, installed the selected skill with all artifacts, and reported security risk assessments from three scanners (Gen, Socket, Snyk). The install completed in approximately 30 seconds.

What the runs tell you

The structural validation confirms the catalog is internally consistent: every skill has the required artifacts. The live install proves the standard npx skills add workflow works correctly with this repo, including security scanning integration. The 41% benchmark failure rate is not an install problem but a real quality signal: nearly half the skills did not pass their own evaluation criteria, which means agents may not benefit from them despite the high production values.

Setup Walkthrough

1. Install a skill using the standard skills CLI:

 npx skills add nvidia/skills

This clones the repo and presents an interactive picker showing all 108 skills grouped by category.

1. To install a specific skill without prompts:

 npx skills add nvidia/skills --skill cuopt-numerical-optimization-api-python --agent claude-code --yes

Replace the skill name with any from the catalog. Use --agent to target Claude Code, Codex, Cursor, or Kiro.

1. Verify the installed skill's signature (requires model-signing pip 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

Post-install: each skill's SKILL.md includes a "Before You Start" section describing prerequisites (NVIDIA hardware, CUDA toolkit, specific product installations). Most skills are non-functional without their corresponding NVIDIA product stack.

Alternatives

1. anthropics/skills (143K stars) -- general-purpose agent skill collection from Anthropic. Much broader applicability across coding tasks, not hardware-specific. No cryptographic signing or governance cards.
2. google/skills (10.9K stars) -- Google's official agent skills for Gemini and Google Cloud. Covers GCP services, not GPU-specific workflows. Lighter on infrastructure verification.
3. microsoft/SkillOpt (2.9K stars) -- automated skill quality optimizer from Microsoft. Complementary rather than competing: optimizes any skill's instructions rather than providing domain-specific skills.