jetson-device-skills

Description

jetson-device-skills is a catalog of Agent Skills for working with a live NVIDIA Jetson device after it has booted. The skills are intended to be present on the Jetson and provide agent-readable instructions plus small helper scripts for diagnostics, memory auditing, runtime selection, model serving, benchmarking, packaging guidance, and related Jetson device workflows.

This repository is device-side. Skills here run on the Jetson, inspect the Jetson, or provide commands that an agent should execute on the Jetson. BSP customization before flashing belongs in the sibling jetson-bsp-skills repository.

This project is currently not accepting contributions.

Included Skills

• jetson-diagnostic captures a read-only device health snapshot, including Jetson identity, memory, GPU, thermal, power, storage, services, and top processes.
• jetson-memory-audit measures DRAM/NvMap usage and verifies before/after memory reclamation with live audit data.
• jetson-headless-mode helps turn a Jetson into a headless edge node by disabling the desktop and safe background services.
• jetson-inference-mem-tune recommends serving runtimes and memory flags for vLLM, SGLang, llama.cpp, and TensorRT Edge-LLM.
• jetson-llm-serve provides Jetson-appropriate vLLM and SGLang serving recipes.
• jetson-llm-benchmark emits structured benchmark metrics for vLLM, llama.cpp, and Ollama paths.
• jetson-package guides Jetson-specific package, wheel, and container choices.
• jetson-speculative-decoding adds Jetson-specific EAGLE-3 or draft-model speculative decoding guidance for vLLM.

Installation

Clone the repository on the Jetson device:

git clone https://github.com/NVIDIA-AI-IOT/jetson-device-skills.git
cd jetson-device-skills

Install the skills into the agent skill directories on the Jetson:

./install.sh

By default, install.sh links the skills into the user-level skill roots for Claude Code, Codex, and Cursor:

• ~/.claude/skills
• ~/.codex/skills
• ~/.agents/skills
• ~/.cursor/skills

You can select specific targets or copy files instead of creating symbolic links:

./install.sh --targets claude,cursor
./install.sh --targets cursor-project --project /path/to/project
./install.sh --copy
./install.sh --force

For NemoClaw/OpenClaw sandboxes:

./install.sh --targets nemoclaw --nemoclaw-sandbox jetson-skills

Restart the agent session after installation so the new skill entries are picked up. If your agent expects a different skill directory, copy or sync the skills/ directory into that location. Keep each skill as a complete directory containing its SKILL.md, scripts/, and references/ content.

Usage

Each skill lives under skills/<skill-name>/ and starts with a SKILL.md file. After install.sh links or copies the skills into the agent's skill directory, agent runtimes such as Cursor, Claude Code, Codex, or NemoClaw/OpenClaw can discover the skills from their frontmatter descriptions and follow the instructions in the selected skill.

Some skills include helper scripts under scripts/, but users normally do not need to call those scripts directly. The agent should invoke the relevant helper when a skill asks for live Jetson data, then use the returned output as the source of truth.

Repository Layout

jetson-device-skills/
├── README.md
├── LICENSE
├── install.sh
├── agents/
└── skills/
    ├── jetson-diagnostic/
    ├── jetson-headless-mode/
    ├── jetson-inference-mem-tune/
    ├── jetson-llm-benchmark/
    ├── jetson-llm-serve/
    ├── jetson-memory-audit/
    ├── jetson-package/
    └── jetson-speculative-decoding/

License

This code is dual-licensed with documentation under the CC-BY-4.0 AND source code under Apache-2.0 license terms. See LICENSE.