Agentic AI — NemoClaw on JetPack 7.2
JetPack 7.2 makes your kit agentic-ready: NVIDIA NemoClaw installs with a single command, and NVIDIA's agent skills automate much of the platform work that used to be manual.
What NemoClaw is
Per NVIDIA: NemoClaw is an open stack/blueprint collection for building autonomous agents — always-on AI systems that reason, plan, and act. It adds privacy and security controls (via OpenShell runtime policy controls) to the OpenClaw agent ecosystem, and packages NVIDIA components such as Nemotron models and NeMo. JetPack 7.2 comes preconfigured with the required dependencies, so no manual environment setup is needed on your kit.
- NemoClaw product page: https://www.nvidia.com/en-us/ai/nemoclaw
- NemoClaw on GitHub: https://github.com/NemoClaw · community examples: https://github.com/nemoclaw-community
Install (single command, official)
On the kit (JetPack 7.2+), run:
curl -fsSL https://www.nvidia.com/nemoclaw.sh | bashSecurity note — read before running: this installs an always-on agent framework. Review what the agent is allowed to access and which credentials it can use before enabling it, prefer scoped/revocable tokens, and use OpenShell's policy controls. Don't leave an agent unattended with access you can't revoke.
After install — where to go next
NVIDIA maintains a Build-a-Claw Resource Hub with installation guidance, cloud trials, and learning resources: https://www.nvidia.com/en-us/ai/build-a-claw
Also useful:
- NVIDIA Deep Learning Institute course: Securing Agents With NemoClaw and OpenShell (see the resource hub)
- NVIDIA Developer Discord —
#nemoclawchannel - Third-party walkthroughs (e.g., Seeed Studio's NemoClaw guide, written for a Jetson Thor robot arm) document post-install flows such as
nemoclaw onboard— treat these as community guidance and follow NVIDIA's hub for the authoritative flow.
Jetson agent skills — automate the platform work
JetPack 7.2 ships agent skills: repeatable, agent-executable workflows for Jetson development. Three categories per NVIDIA:
| Skill category | What it automates |
|---|---|
| Jetson Linux customization | Building/customizing a BSP for custom carrier boards — I/O config, clocks, fan control, power profiles |
| Memory optimization | Auditing bootloader carveouts, kernel reservations, and user-space memory to fit more capable workloads in less memory |
| Model benchmarking | Finding the optimal model configuration and diagnostics for your device |
More agent skills in the ecosystem:
- Jetson device-side skills · Jetson BSP skills
- DeepStream Coding Agent — agent-assisted vision pipeline building (see our DeepStream tutorial)
- Metropolis VSS blueprint skills — video search and summarization workflows
Practical notes for the AGX Orin kit
- Always-on agents need dedicated compute — that's the point of running them on a kit rather than a laptop that sleeps; plan power and thermals accordingly (see the power-mode notes in Troubleshooting).
- Model choice matters for memory — local models on Orin run well within 64GB, but always-on agents accumulate context. See Memory Efficiency for the levers, and Local LLM Inference for on-device model performance.
- This space moves fast. Treat the commands above as the current official path; check the resource hub for updates before scripting deployments.
Sources
- NVIDIA Technical Blog — JetPack 7.2 agentic AI (install command, agent skills, release features) (checked 2026-09-24)
- NVIDIA NemoClaw product page (checked 2026-09-24)
- JetPack 7.2.1 downloads page (checked 2026-09-24)
Status: draft, pending review by cheny. Grounded in NVIDIA's official documentation as of the date listed; not yet verified on physical hardware by Juxi Technology.
NVIDIA® and Jetson™ are trademarks of NVIDIA Corporation. This page is published by Juxi Technology and is not an NVIDIA publication.