As AI agents move from experiments to production, teams need verifiable, reusable instruction sets that make agents call platform APIs correctly and safely. This repository centralizes NVIDIA-published agent skills—small, opinionated instruction bundles that teach an agent how to use a specific NVIDIA library, service, or workflow in a repeatable way.
What Sets It Apart
- NVIDIA verification and detached signatures: each skill ships with a detached OMS signature and a verifiable trust anchor so consumers can detect tampering — this supports governance and supply-chain integrity.
- Evaluation artifacts included: published skills include benchmark/evaluation JSON and Tier-3 datasets where applicable, enabling measurable uplift and reproducible verification before adoption.
- Spec-compatible, agent-agnostic packaging: skills follow the Agent Skills spec so they can be installed into multiple agent clients (examples in the catalog include integrations for common agent runtimes), reducing bespoke integration work.
- Focused on platform and workflow correctness: skill content emphasizes correct usage patterns for CUDA-X libraries, cuOpt, NeMo blueprints, video/vision pipelines, and deployment blueprints rather than providing new models or runtimes.
Who It's For and Tradeoffs
Great fit if you run or build agent-driven ML workflows, MLOps pipelines, or platform integrations that rely on NVIDIA software and need verifiable, reusable agent behaviors. It helps reduce integration bugs and provides reproducible evaluation artifacts for governance.
Look elsewhere if you need a runtime agent framework, a complete end-user chatbot UI, or vendor-agnostic skills: the catalog is curated around NVIDIA software and assumes an agent runtime that implements the Agent Skills spec and can consume the skill artifacts. Adopting skills still requires runtime support, infra, and engineering to integrate them into your product pipeline.
