NVIDIA Agent Skills

NVIDIA/skills is NVIDIA's official catalog of verified Agent Skills for AI coding agents. It teaches agents how to work with NVIDIA software and platform workflows across CUDA-X libraries, RAG Blueprint, NemoClaw and OpenClaw, NeMo, TAO, cuOpt, Omniverse, Physical AI, video search, simulation, robotics, vision AI, medical AI, and GPU-accelerated data science.

This is not a generic prompt collection. The repository describes daily sync from product repositories, signed skill artifacts, skill cards, benchmark reports, evaluation datasets, security scanning, and marketplace syndication for Agent Skills-compatible clients.

Knowledge Freshness

The repository is active and skills are mirrored from product teams into the catalog by an automated sync pipeline. NVIDIA notes that skills are added continuously. GPU software, CUDA libraries, RAG services, NIM models, NeMo recipes, TAO workflows, NemoClaw sandbox behavior, and deployment guidance can change quickly.

Use the skills for product-specific workflow routing and source-backed guardrails, then verify commands, versions, driver requirements, model names, container images, and deployment assumptions against the current upstream product repository and your target environment.

Retrieval Sources

This listing is grounded in:

• The NVIDIA/skills repository README.
• The skills.sh.json grouping metadata.
• The Codex plugin manifest for nvidia-skills.
• Representative SKILL.md files for NemoClaw/OpenClaw onboarding, RAG Blueprint operations, and cuOpt Python optimization.
• The repository license text.
• Current GitHub repository metadata from the official NVIDIA organization.

Core Workflow

List available NVIDIA skills:

npx skills add nvidia/skills --list

Install one skill for Codex:

npx skills add nvidia/skills --skill rag-blueprint --agent codex

Install NemoClaw/OpenClaw onboarding guidance for Claude Code:

npx skills add nvidia/skills --skill nemoclaw-user-get-started --agent claude-code

Install the NVIDIA plugin marketplace bundle for Codex:

codex plugin marketplace add NVIDIA/skills

Capability Scope

The catalog currently groups skills across these areas:

The plugin manifest frames the bundle as guided help for NVIDIA AI, GPU, robotics, simulation, and 3D workflows. It also includes example prompts for CUDA-X research, physical-world video to 3D neural scene workflows, and secure OpenClaw deployment.

Production Rules

Treat these skills as high-powered operational guidance:

• Start with a narrow product skill instead of applying broad NVIDIA advice to every GPU problem.
• Verify hardware, driver, CUDA, container, Kubernetes, and API-key prerequisites before running deployment or training commands.
• For NemoClaw/OpenClaw, review installer behavior, Docker access, inference provider registration, web search, messaging tokens, network policy tiers, dashboard URLs, and sandbox rollback behavior before accepting prompts.
• For RAG Blueprint work, detect the active deployment mode before changing Docker Compose, Helm, vector database, NIM, guardrails, observability, or ingestion configuration.
• For GPU training and inference, estimate compute cost, data movement, model artifacts, and rollback path before launching long-running jobs.
• For signed skills, verify detached OMS signatures against the NVIDIA trust anchor when supply-chain integrity matters.
• Keep generated data, medical artifacts, videos, embeddings, model weights, logs, and deployment output out of public prompts unless explicitly approved.

Use Cases

• Ask Codex to deploy or troubleshoot NVIDIA RAG Blueprint using the right Docker, Helm, NIM, vector database, and observability paths.
• Use Claude Code to install NemoClaw, configure an OpenClaw sandbox, and review security policy choices before running the first agent prompt.
• Ask Cursor to model and solve an LP, MILP, or QP with the cuOpt Python API.
• Use an agent to route Physical AI or Omniverse workflows from CAD to simulation-ready assets.
• Give an AI assistant source-backed NVIDIA guidance for TAO, DeepStream, NeMo, Megatron-Core, video search, CUDA-Q, Holoscan, Dynamo, or cuDF work.

Source Review

• The README describes NVIDIA Agent Skills as official, NVIDIA-verified skills for AI agents.
• The README states that the catalog teaches agents how to use NVIDIA software including CUDA-X libraries, AI Blueprints, and platform tools.
• The catalog documents install flows for the skills CLI, Claude Code, Codex, Cursor, and Kiro CLI.
• The README states that every published skill ships with SKILL.md, skill-card.md, skill.oms.sig, a Tier-3 evaluation dataset, and BENCHMARK.md.
• The Codex plugin manifest describes guided help for NVIDIA AI, GPU, robotics, simulation, and 3D workflows, with Interactive and Write capabilities.
• The NemoClaw quickstart skill covers installing NemoClaw, launching a sandbox, registering inference, applying policies, and running OpenClaw.
• The RAG Blueprint skill covers deployment, configuration, troubleshooting, shutdown, Docker/Compose, Kubernetes/Helm, MCP configuration, observability, guardrails, ingestion, and model infrastructure.
• The license file identifies the repository as Apache-2.0 for code and configuration plus CC-BY-4.0 for documentation.

Duplicate Review

Checked current content/skills/, content/tools/, content/mcp/, open pull requests, and repository-wide content for NVIDIA Agent Skills, NVIDIA/skills, nvidia-skills, NemoClaw, OpenClaw, rag-blueprint, cuopt-numerical-optimization-api-python, and related NVIDIA skill phrases. No dedicated NVIDIA Agent Skills entry, exact source URL duplicate, target file, or open duplicate PR was found.

Disclosure

Editorial listing. No paid placement or affiliate link is used. The upstream repository is published by the official NVIDIA GitHub organization. Its license states Apache-2.0 for code and configuration plus CC-BY-4.0 for documentation.