One email each morning — yesterday's AI, sortedGet it in your inbox

METAL LAB

nvidia/skills

3,032PythonApache-2.0

A ready-made instruction library that teaches coding AI agents how to use NVIDIA tools

nvidia/skills is a catalog of Agent Skills — instruction packages you install into coding agents like Claude Code and Codex so they know how to operate NVIDIA products such as CUDA, robotics simulation, and RAG pipelines. Skills are actually maintained in their own product repos and mirrored into this catalog through a daily automated sync. Every published skill carries an NVIDIA signature so users can verify it hasn't been tampered with.

What it does

  1. The catalog organizes dozens of NVIDIA products into skills, covering things like cuOpt (optimization), DeepStream (video analytics), Jetson (embedded boards), TAO Toolkit (vision model fine-tuning), and the NeMo family (LLM training).
  2. A single command, npx skills add nvidia/skills, lets you pick a skill and install it into Claude Code, Codex, Cursor, and other supported agents.
  3. Every skill is required to ship SKILL.md (instructions), skill-card.md (identity and governance metadata), skill.oms.sig (a signature file), and an evaluation dataset — skills missing these are automatically dropped from the catalog before publishing.
  4. Anyone can verify a skill hasn't been modified since NVIDIA signed it, using the provided nv-agent-root-cert.pem trust certificate.
  5. The repository has 3032 stars, and its skills are kept in their upstream product repositories, mirrored into this catalog daily via an automated sync pipeline.

Why it matters

It addresses the problem of AI coding agents fumbling with unfamiliar hardware SDKs or specialized software by packaging verified, standardized instructions instead of relying on ad hoc prompting. Developers can install a vetted skill and hand off CUDA or robotics simulation tasks to their agent right away, with a built-in way to confirm the instructions are authentic.

Terms in this repo

  • Agent Skills · Portable instruction packages that teach an AI agent how to correctly use a specific tool or piece of software
  • SKILL.md · The file containing the actual instructions an agent reads and follows
  • OMS signature · A digital signature format used to prove files haven't been altered since signing
  • RAG · A method that generates answers grounded in documents retrieved by search
  • CUDA · NVIDIA's platform for writing parallel programs that run on its GPUs

Repository description (English)

Agent Skills for NVIDIA products — install into Claude Code, Codex, and other coding agents to run Physical AI, robotics, simulation, CUDA, and RAG workflows end to end.

Open on GitHubProject site

Coverage

Trending repos

All repos →

Latest from METAL LAB