# Jetson Agent Skills

> Source: https://aiwiki.ai/wiki/jetson_agent_skills
> Updated: 2026-10-10
> Fact-checked: 2026-10-10
> Categories: AI Agents, Developer Tools, Edge computing, NVIDIA
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Jetson Agent Skills." aiwiki.ai, 10 Oct 2026. https://aiwiki.ai/wiki/jetson_agent_skills
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Jetson Agent Skills are [NVIDIA](https://aiwiki.ai/wiki/nvidia) instructions and helper files that let coding agents carry out device-specific development workflows on [NVIDIA Jetson](https://aiwiki.ai/wiki/nvidia_jetson). The collection has two families: Device Skills for a running Jetson and BSP Skills for preparing and customizing its Linux Board Support Package on a host workstation.[1]

The coding agent's tools operate against the development environment; its language model can run locally or in the cloud. Installing skills therefore does not by itself make the agent's model run on the Jetson.[8]

## Device Skills and BSP Skills

| Family | Execution environment | Main purpose |
|---|---|---|
| Jetson Device Skills | A booted Jetson | Application setup, diagnostics, inference, memory use, and video workflows.[1] |
| Jetson BSP Skills | A host workstation with a `Linux_for_Tegra` tree | Hardware configuration, BSP customization, building, flashing, and validation.[1] |

An [agent skill](https://aiwiki.ai/wiki/agent_skills) is a directory with a `SKILL.md` instruction file, a description that helps the agent select it, and optional helpers. Jetson skills combine documented procedures with facts from the actual target, such as its installed software, memory, and power mode.[1]

## Device skill catalog

The device repository includes the following workflows.[2]

| Skill | Purpose |
|---|---|
| `jetson-diagnostic` | Inspect identity, memory, GPU activity, temperature, power, storage, and services.[2] |
| `jetson-memory-audit` | Measure memory use and check changes after reclamation.[2] |
| `jetson-headless-mode` | Plan and apply changes to desktop and background services.[2] |
| `jetson-inference-mem-tune` | Select inference runtimes and memory settings.[2] |
| `jetson-llm-serve` | Configure model serving.[2] |
| `jetson-llm-benchmark` | Measure LLM inference performance.[2] |
| `jetson-package` | Identify compatible packages, wheels, and containers.[2] |
| `jetson-speculative-decoding` | Explore speculative decoding configurations.[2] |
| `jetson-video-setup` | Prepare and verify codec software.[2] |
| `jetson-video-capability` | Check working encode and decode capabilities.[2] |
| `jetson-video-recipe` | Produce a codec configuration plan.[2] |
| `jetson-video-benchmark` | Measure encoding and decoding performance.[2] |
| `jetson-video-pipeline` | Execute and verify codec pipeline stages.[2] |

The repository installer links or copies complete skill directories into the locations used by [Claude Code](https://aiwiki.ai/wiki/claude_code), [Codex](https://aiwiki.ai/wiki/openai_codex), Cursor, or supported [OpenClaw](https://aiwiki.ai/wiki/openclaw) sandboxes. The agent session must restart after installation to discover them.[2]

## Diagnostics and package compatibility

The diagnostic skill collects a read-only health snapshot. Its helpers return JSON or a compact memory summary, and the agent must report values obtained from that output. Diagnosis does not authorize changing power settings, stopping services, installing packages, or starting a model server.[4]

A sandbox may expose the skill files without exposing Jetson hardware. Missing host paths or tools can prevent measurements, and NvMap debug information may require root access. The diagnostic instructions require agents to preserve missing or unknown fields rather than invent substitute device readings.[4]

The package skill checks the Jetson generation, JetPack release, and CUDA build target before recommending artifacts. ARM64 compatibility alone does not establish that a wheel contains the required GPU kernels. Its helper produces compatibility hints and catalog locations; it does not install packages or pull containers.[6]

Package selection also does not verify that a particular model checkpoint fits the device's memory. Container tags change, so placeholders in documentation require a current catalog lookup rather than literal use as an installation target.[6]

## BSP preparation and deployment

BSP Skills organize development around an active target profile and four stages: setup, customization, build, and deployment. Setup registers or downloads inputs and prepares image and source workspaces. Customization covers areas such as pinmux, USB, PCIe, cameras, clocks, power profiles, and reserved memory. Build regenerates affected artifacts; deployment promotes changes into the image, flashes the device, and validates it.[3]

The bundle supplies instructions, templates, and references, not a BSP image or board documents. Its README demonstrates a Claude Code workflow starting with `jetson-quick-start`. Engineers must review generated plans, commands, and diffs. Flashing can erase storage or leave the target unbootable, so backups and checks of the target, release, and hardware remain necessary.[3]

## Approval and reversible memory changes

The headless-mode skill illustrates the separation between a proposed change and its execution. It starts from a current memory audit, produces a plan, and defaults to a dry run. Mutations require explicit approval and `sudo`; the apply helper needs `--apply` to make changes.[5]

The workflow is inappropriate when the user needs a local desktop, kiosk display, or graphical session. It excludes boot configuration and device-tree edits, retains specified camera and recovery services, and records reversal commands. Estimated memory savings are upper bounds; a fresh audit measures the actual before-and-after difference.[5]

## Video workflows

NVIDIA's 11 August 2026 JetPack 7.2.1 announcement describes foundational video skills above Video Codec SDK and PyNvVideoCodec 2.2. They inspect the target, choose a supported codec path, generate configurations, execute measurements, and return warnings and reproducible artifacts. Benchmarks can include throughput, latency, utilization, bitrate, and quality.[7]

These skills configure and verify codec stages. The announcement explicitly excludes additional GStreamer, V4L2, AI-model, or application-level pipeline-building skills for that release. A coding assistant can add preprocessing, inference, or output handling around the codec stages, but those application choices are separate from the video skill's scope.[7]

## Development context

NVIDIA described device-side and BSP-side agent skills in its 1 June 2026 JetPack 7.2 announcement, including Linux customization, memory optimization, and model benchmarking.[9]

The Jetson AI Lab walkthrough keeps the developer responsible for approvals and verification on real hardware. Its development-time graphic is illustrative, not a measured speedup for Jetson Agent Skills. It also distinguishes an interactive SSH session from a persistent deployment: unattended services need explicit configuration and testing.[8]

## References

1. NVIDIA Jetson AI Lab. [Jetson Agent Skills](https://www.jetson-ai-lab.com/tutorials/jetson-agent-skills/), accessed 10 October 2026.
2. NVIDIA-AI-IOT. [Jetson Device Skills repository README](https://github.com/NVIDIA-AI-IOT/jetson-device-skills), accessed 10 October 2026.
3. NVIDIA-AI-IOT. [Jetson BSP Skills repository README](https://github.com/NVIDIA-AI-IOT/jetson-bsp-skills), accessed 10 October 2026.
4. NVIDIA-AI-IOT. [Jetson Diagnostic skill](https://github.com/NVIDIA-AI-IOT/jetson-device-skills/blob/main/skills/jetson-diagnostic/SKILL.md), accessed 10 October 2026.
5. NVIDIA-AI-IOT. [Jetson Headless Mode skill](https://github.com/NVIDIA-AI-IOT/jetson-device-skills/blob/main/skills/jetson-headless-mode/SKILL.md), accessed 10 October 2026.
6. NVIDIA-AI-IOT. [Jetson Package and Environment skill](https://github.com/NVIDIA-AI-IOT/jetson-device-skills/blob/main/skills/jetson-package/SKILL.md), accessed 10 October 2026.
7. Shashank Maheshwari and colleagues, NVIDIA Technical Blog. [NVIDIA JetPack 7.2.1 Adds Agentic Video Skills and T3000 Emulation](https://developer.nvidia.com/blog/nvidia-jetpack-7-2-1-adds-agentic-video-skills-and-t3000-emulation/), 11 August 2026.
8. NVIDIA Jetson AI Lab. [A New Way to Build Edge AI](https://www.jetson-ai-lab.com/tutorials/ai-assisted-development-on-jetson/), accessed 10 October 2026.
9. Peilun Tsai and Shashank Maheshwari, NVIDIA Technical Blog. [Deploy Agentic-Ready AI at the Edge with Memory Efficiency in NVIDIA JetPack 7.2](https://developer.nvidia.com/blog/deploy-agentic-ready-ai-at-the-edge-with-memory-efficiency-in-nvidia-jetpack-7-2/), 1 June 2026.