NVIDIA Agent Skills is a catalog of portable instruction sets that teach AI coding agents how to use NVIDIA software optimally, installable into Claude Code, Codex, and other coding agents.
The tool addresses the problem of AI agents lacking domain knowledge about NVIDIA's specialized software ecosystem. Rather than embedding this knowledge directly into agent models, the project distributes skills as modular, portable instruction sets that agents can load on demand. Skills teach agents how to use Physical AI and robotics workflows, simulation tools, CUDA-X libraries, RAG systems, and platform tools. The skills themselves are maintained in their respective product repositories and mirrored into this catalog daily through an automated sync pipeline, ensuring they stay current with upstream changes.
Developers should adopt this if they are building with AI agents and need those agents to work effectively with NVIDIA's software stack. The approach suits teams using Claude Code, Codex, Cursor, Snowflake CoCo, or Kiro who want their agents to understand CUDA, cuOpt, simulation, or robotics APIs without manual prompting. Installation uses the standard skills CLI, which handles agent-specific targeting and keeps installed skills synchronized with upstream updates. The project explicitly positions itself as NVIDIA-verified capability governance for agents, distinguishing it from unvetted or community-maintained skill collections.
The project maintains an automated daily sync pipeline that mirrors skills from their source repositories into the catalog, ensuring the collection stays synchronized with upstream development. The team actively consolidates and evolves the skill catalog, removing or merging skills as the ecosystem matures and flagging these changes to users during updates. Skills are being added continuously, with the roadmap publicly visible and contributions welcomed from the community.
