OpenClawDocs
On this page

On this page

View as MarkdownView this page as plain textOpen in ChatGPTAsk questions about this pageOpen in ClaudeAsk questions about this page

Hosting

NVIDIA Jetson

Run OpenClaw on a Jetson using the standard Linux installation. Start with a cloud model provider; running a local model server is a separate setup with its own memory and GPU requirements.

Prepare the Jetson

  1. Install a JetPack release supported by your module and carrier board. Follow NVIDIA's JetPack instructions or SDK Manager for flashing and storage options. JetPack versions support different boards; this guide does not certify every Jetson module or image.
  2. Use a 64-bit Linux image. Check uname -m; it should report aarch64.
  3. Check the Node.js requirements. Use a supported Node release and linked SQLite library, rather than assuming the image's distro Node is compatible. The OpenClaw installer checks the runtime and installs Node when needed.
  4. Use storage, cooling, and a power supply appropriate for your board. Consider an SSD for an always-on host when the board supports it.

Install and connect

Follow Getting Started for installation, onboarding, and verification. For a headless Jetson, the Linux CLI and SSH instructions cover service installation and an SSH tunnel to the Control UI. Keep the Gateway on loopback when using that tunnel.

Linux onboarding attempts to enable systemd lingering so the user service can survive logout. If that step fails or the Gateway stops after logout, follow the Gateway service guidance to check lingering and the user session bus. This is standard Linux service behavior.

Jetson diagnostics

On JetPack images that include NVIDIA's tools:

bash
sudo nvpmodel -qtegrastats

nvpmodel -q displays the current power mode. tegrastats reports resource use and temperatures; press Ctrl+C to stop it. Power-mode IDs and limits vary by module and image. Follow the power and performance documentation for your JetPack release before changing modes or clocks; see NVIDIA's Jetson Linux Developer Guide.

If the host runs out of memory, inspect available RAM and swap with free -h and swapon --show. Compressed RAM swap (zram) shares physical memory with the rest of the system. Size a local-model workload for the memory actually available to it; an OpenClaw installation does not validate GPU inference compatibility or model capacity.

Next steps

  • Local models — connect a separately configured model server
  • Linux server — remote access and server operation
  • Backups — preserve configuration, state, and workspace
  • Updating — keep OpenClaw up to date
Was this useful?