OpenClaw on NVIDIA DGX Spark: setup with Claude Code

This tutorial builds an OpenClaw AI agent on NVIDIA DGX Spark using an existing Ollama installation. Claude Code handles installation and infrastructure troubleshooting, with Opus selected at high effort. The presenter then configures OpenClaw's identity and memory files, pairs a Telegram bot, and accesses the dashboard from another device through an SSH tunnel.

Two practical exercises cover a scheduled morning news brief and an Australian fuel-price tracker. The dashboard exposes tool calls and file writes so the presenter can check whether the agent actually fetched sources, saved a skill, and created scheduled jobs. The first news brief invents stories and fails to save the expected automation.

Troubleshooting reveals a local LLM configuration mismatch: Ollama gives Gemma a 32k token context window while OpenClaw expects nearly 200k. After correcting that, Gemma hits the task timeout, prompting a switch to a Qwen model. These are results from the demonstrated setup, rather than universal model limits.

Although the presenter describes local inference as capable of offline operation, these examples use external websites, a fuel API, and Telegram. The closing discussion covers prompt injection, restricted permissions, human review, and NemoClaw. The presenter describes NemoClaw as a deny-by-default security layer, but does not install it in this walkthrough.