Jetson Agent Skills : AI-Assisted Workflows for Device & BSP Customization

Robotics & Edge Computing Jetson Systems Jetson Thor
, , , ,
1

Hi Jetson Developers 👋,

We are excited to share two open-source Jetson Agent Skills repositories that bring AI-assisted workflows to NVIDIA Jetson development.

These skills help AI coding assistants such as Claude Code, OpenAI Codex, Cursor, and NemoClaw/OpenClaw perform Jetson-specific tasks more reliably. Instead of manually searching documentation, running many diagnostic commands, or editing BSP files from scratch, you can describe the task in natural language and let the agent follow guided Jetson workflows.

Jetson Device Skills

Repository: jetson-device-skills

Jetson Device Skills run directly on your Jetson device and help with post-boot device management, diagnostics, performance tuning, and AI workload setup.

Available skills include:

  • jetson-diagnostic - Collects a full device health snapshot, including memory, GPU, thermal, power, storage, and process information.
  • jetson-memory-audit - Reviews DRAM and NvMap usage and helps verify memory reclamation.
  • jetson-headless-mode - Converts a Jetson device into a headless edge node by disabling desktop and selected services.
  • jetson-inference-mem-tune - Recommends runtime and memory settings for vLLM, SGLang, llama.cpp, and TensorRT Edge-LLM.
  • jetson-llm-serve - Provides Jetson-optimized serving recipes for vLLM and SGLang.
  • jetson-llm-benchmark - Captures structured benchmark metrics for vLLM, llama.cpp, and Ollama.
  • jetson-package - Provides Jetson-specific package, wheel, and container guidance.
  • jetson-speculative-decoding - Helps configure EAGLE-3 and draft-model speculative decoding for vLLM.

🚀 Quick Start

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

Jetson BSP Skills

Repository: jetson-bsp-skills

Jetson BSP Skills run on your host machine and guide the BSP customization lifecycle, from workspace setup through flashing and validation.

Supported workflow areas include:

  • Setup - Initialize the target, download the BSP, extract images, and prepare sources.
  • Customize - Update pinmux, USB, PCIe, UPHY, MGBE, camera, clocks, fan, nvpmodel, and memory-related configuration.
  • Build - Rebuild DTBs, kernel modules, and other generated artifacts after changes.
  • Deploy - Promote workspace changes, flash the device, and validate the result.

For example, you can ask your agent:

Configure PCIe controller 5 for Gen4 x4 endpoint mode.

The BSP skill guides the agent through the relevant DTSI changes, overlay structure, build steps, and validation workflow.

Quick Start

git clone https://github.com/NVIDIA-AI-IOT/jetson-bsp-skills.git
cd jetson-bsp-skills
./setup.sh --workspace ~/my-bsp-workspace
cd ~/my-bsp-workspace
claude

You can use Claude Code or another supported agent runtime.

Tutorials

Step-by-step walkthroughs are available on Jetson AI Lab:

Requirements

  • Agent runtime: Claude Code, OpenAI Codex, Cursor, or NemoClaw/OpenClaw
  • Supported Jetson platform: Jetson Orin Nano, Orin NX, AGX Orin, or Thor

Your Feedback is always welcome

We are actively developing additional Jetson skills and would like feedback from the community.

Please share:

  • Which skills are most useful for your workflow?
  • What Jetson tasks would you like to see covered next?
  • Any issues, bugs, or suggestions from trying the repositories.

Reply in this thread or open an Issue or Discussion on the GitHub repositories. Your feedback will help shape the next set of Jetson Agent Skills.

4

Thanks for releasing these — the split between on-device skills and host-side BSP skills is clear.

For my workflow, jetson-diagnostic and jetson-memory-audit look the most immediately useful, especially when chasing thermal/memory issues after long inference runs. On the BSP side, camera and pinmux customization is where I usually burn the most time, so guided DTSI/overlay steps there would help a lot.

One question: for Jetson Device Skills, is the intended path to run the agent runtime directly on the Jetson, or drive it from a host machine over SSH? Also curious whether Orin Nano 8GB is a realistic target for the LLM serve/benchmark skills, or if those assume more memory (Orin NX / AGX / Thor).

Will try the device-skills install on an Orin and report back if I hit anything odd.