NVIDIA-Verified Agent Skills

Agent skills are portable instruction sets that extend what an AI agent can do. They can teach an agent how to use NVIDIA CUDA-X libraries, AI Blueprints, Omniverse and Physical AI workflows, NeMo training and inference tools, and other platform components correctly.

As skills become a reusable capability layer, teams need more than runtime guardrails. They need to know what capability entered the workflow, where it came from, whether it was reviewed for common risks, and whether it changed after publication. NVIDIA-verified skills provide that capability governance for the skill layer.

For the current catalog, use skills.sh or NVIDIA Build as the source of truth. Install from the catalog with:

$
npx skills add nvidia/skills

Works Across Agents

NVIDIA skills build on the open Agent Skills specification, so they are designed to work across compatible agent clients rather than being tied to one runtime.

Vercel’s skills CLI is the delivery vehicle for bringing NVIDIA skills to agents beyond Claude Code and Codex, including OpenClaw, Kiro, Aider, Augment, and other compatible agents. See the CLI’s supported agents list for the current targets. Developers can use the CLI to install from the NVIDIA catalog today:

$
npx skills add nvidia/skills --agent codex

Marketplace and hub discovery are additional distribution paths. Developers will be able to discover the NVIDIA skills plugin directly through the Claude Code and Codex marketplaces. NVIDIA skills are also planned for Hermes Skillhub and Clawhub as those partner channels come online.

What Skills Give Your Agent

Install a skill once, and the agent can load product-specific procedures, references, scripts, and safety boundaries when a matching task comes up. With the right NVIDIA skills installed, an agent can:

• Formulate and solve routing, scheduling, linear programming, and quadratic programming tasks, for example with cuOpt skills.
• Deploy, configure, evaluate, and troubleshoot RAG, AI-Q, NemoClaw, and agent sandbox workflows.
• Bring up inference and serving workflows, for example with Dynamo and NeMo Platform skills.
• Build, validate, and tune distributed model training, for example with NeMo, Megatron-Core, DALI, and Nemotron skills.
• Accelerate data and scientific workloads, for example with cuDF, cuPyNumeric, CUDA-Q, Earth2Studio, PhysicsNeMo, and TileGym skills.
• Build vision, video, Physical AI, and Omniverse USD workflows, for example with DeepStream, Video Search and Summarization, CAD-to-SimReady conversion, realtime viewing, USD performance tuning, and neural reconstruction skills.

How Skills Are Discovered

The catalogs group skills by the kind of work the user is trying to do, while this repository preserves product ownership and source provenance. Start with the workflow when you know the task; use skills.sh or NVIDIA Build when you need the current skill list.

Trust Controls

NVIDIA-verified skills are also release artifacts. Before a skill is published, the catalog expects the same discipline we expect from deployable software: security review, provenance, ownership, clear use boundaries, and verification instructions.

Verified means the skill is cataloged, scanned, signed, and documented with a skill card:

• Cataloged from the product team or source repository that owns the capability.
• Scanned before publication for software risks and agent-native risks such as hidden instructions, prompt injection, trigger abuse, excessive agency, and mismatches between declared purpose and bundled behavior.
• Signed with a detached skill.oms.sig so users can verify that the downloaded skill is authentic and unchanged.
• Documented with a skill card that records what the skill does, who owns it, how it is licensed, what it depends on, and what limitations or risks users should understand.

Runtime controls govern what an agent can do during execution. Verified skills govern which capabilities enter the workflow in the first place.