NVIDIA Agent Skills
Official NVIDIA-verified skills for coding agents to run Physical AI, robotics, simulation, CUDA, and RAG workflows end-to-end.
Evidence shows security policy (SECURITY.md) and DCO check, but no specific implementation of least privilege, user confirmation, data flow transparency. Dependency security: workflow pins third-party action to a commit, but no dependency audit. External effects: skills may execute system commands, but no explicit restrictions. Rollback: CLI provides update and remove, but no explicit rollback mechanism. Source attribution: clearly NVIDIA official, with copyright notices. Deductions: lack of concrete security mechanism evidence.
Self-consistency: README and tests show consistent skill structure, but not fully verified. Dependency availability: relies on external CLI and network resources, no offline guarantee. Failure messages: tests cover error handling, but not comprehensive. Deductions: external dependencies, limited failure message coverage.
Audience and scenarios: README clearly targets multiple coding agents and NVIDIA products, broad scenarios. Capability boundaries: skill catalog clear, but individual skill boundaries not explicit. Trigger precision: skills installed via CLI, but trigger conditions not explicit. Environment fit: supports multiple agents and platforms, but not fully tested. Deductions: capability boundaries and trigger precision underdescribed.
Information architecture: README and skill catalog well organized. Install notes detailed, multiple installation methods. Naming stability: skill names may evolve, no stability guarantee. Examples and FAQ: README provides examples, FAQ missing. Known limitations: only CLI version requirement mentioned, not comprehensive. License clear: Apache-2.0 and CC-BY-4.0. Versioning/changelog: no explicit version numbers or changelog. Maintenance responsibility: NVIDIA maintains, automated sync. Deductions: naming stability, known limitations, versioning/changelog insufficient.
Output usability: skills provide structured output (e.g., JSON), tests verify formats. Marginal value: skills cover multiple NVIDIA products, provide expert guidance, high value. Cost-benefit: low installation and usage cost, but no performance or resource consumption data. Deductions: cost-benefit lacks data support.
Claim traceability: README claims NVIDIA-verified, but no verification evidence. Cross-source corroboration: skill content may come from product docs, but not explicitly cited. Fact-inference separation: skills may mix facts and guidance, but not clearly separated. Deductions: lack of verification evidence and source citations.
- Skills may execute system-level commands; test in isolated environments.
- Dependencies on external CLI and network may not work offline.
- Skill content may not be up-to-date; verify compatibility with current product versions.
What does this agent do, and when should you use it?
NVIDIA Agent Skills is an official catalog of agent skills verified by NVIDIA, designed to enhance coding agents like Claude Code and Codex with expertise in NVIDIA software. The repository hosts hundreds of workflow skills spanning Physical AI, robotics, simulation, CUDA libraries, RAG blueprints, and platform tools. Skills are maintained in their respective product repos and mirrored daily via an automated sync pipeline. Each skill includes SKILL.md instructions, a skill card, an OMS signature, and evaluation datasets, ensuring verifiability and supply-chain security. Installation is managed via the skills CLI, which supports selecting specific skills and target agents, with both interactive and non-interactive modes. The catalog also supports updating and listing skills for continuous integration of NVIDIA's latest capabilities.
This repository acts as a catalog of portable skill instructions that agents load and execute on demand. Skills guide agents through specific NVIDIA workflows, such as solving linear programming problems with cuOpt, fine-tuning vision models with TAO, or deploying RAG blueprints. Upon installation, agents automatically invoke skills when encountering relevant tasks; for example, asking 'solve a linear programming problem with cuOpt' guides the agent through the cuOpt Python API. Skills are sourced from product repos (e.g., cuOpt, TAO, RAG Blueprint) and synced daily, ensuring up-to-date and signed content. Users install via npx skills add nvidia/skills, specify a skill with --skill, or target a specific agent with --agent (e.g., claude-code or codex).
- Developers using Claude Code or Codex who need guidance to leverage NVIDIA cuOpt for GPU-accelerated optimization, like vehicle routing problems.
- Robotics engineers using Isaac Sim for simulation, requiring agent assistance to build Holoscan applications or set up sensor bridge workflows.
- Data scientists looking to fine-tune vision models with TAO Toolkit, wanting the agent to autonomously execute training pipelines based on official NVIDIA guidance.
- AI application developers deploying RAG systems, using skills to configure Docker Compose or Helm, troubleshoot, and optimize performance.
- HPC users migrating NumPy code to cuPyNumeric for multi-node GPU acceleration, with skills to guide installation and migration steps.
What are this agent's strengths and limitations?
- Official NVIDIA verification ensures accuracy and alignment with NVIDIA product best practices
- Broad coverage across product domains (Physical AI, robotics, CUDA, RAG) for diverse needs
- Skills are signed with OMS signatures and include evaluation data for verifiable supply-chain security
- Supports multiple agents (Claude Code, Codex, Cursor, etc.) for flexibility
- Continuously updated with daily sync for latest skills
- Dependent on NVIDIA ecosystem and the skills CLI, requiring Node.js 22 environment
- Skills are primarily for NVIDIA products, limiting usability for non-NVIDIA platforms
- Some skills require specific hardware (e.g., GPU) or additional setup (e.g., Docker Compose), increasing deployment overhead
- Currently supports only certain agents (Claude Code, Codex), others require adaptation
- Installation and updates depend on network access to npm and GitHub, which may be restricted
How do you install or deploy this agent?
Requires Node.js. Install the skills CLI (v1.5.16 or later). Run npx skills add nvidia/skills for interactive installation, selecting desired skills and destination. For non-interactive installation, use npx skills add nvidia/skills --skill cuopt-numerical-optimization-api --yes. Specify the target agent with --agent (e.g., --agent claude-code or --agent codex).
How do you use this agent?
After installation, agents like Claude Code automatically load skills when relevant tasks arise. For example, ask 'solve a linear programming problem with cuOpt' and the agent follows the skill's guidance to use the cuOpt Python API. In Claude Code, run /reload-skills to load new skills in the current session. Use npx skills list to see installed skills, npx skills update to update them, and npx skills check to preview outdated ones.
FAQ
Do I have to pay for these skills?
Can these skills run on non-NVIDIA hardware?
How do I verify a skill's signature?
model_signing verify certificate SKILL_DIR --signature SKILL_DIR/skill.oms.sig --certificate_chain nv-agent-root-cert.pem --ignore_unsigned_files to ensure the skill hasn't been tampered with.What happens to old versions after a skill update?
npx skills update interactively updates skills and flags removed or merged ones, offering to remove stale local copies.