Punching Through NVIDIA NemoClaw's Sandbox to Hit Local vLLM on RTX 5090

The author documents an experimental method to route NVIDIA NemoClaw's sandboxed AI agent inference to a local vLLM instance running Nemotron 9B on an RTX 5090, bypassing three layers of network isolation (host iptables, k3s network policy with TCP relay, and sandbox iptables injection). A tool call gateway is also needed to translate Nemotron's raw tool call output into OpenAI-compatible structured format.