What Is the Best Way to Run an OpenClaw Agent Without Sending Prompts to the Cloud?
Last updated: 4/28/2026
Summary: NemoClaw runs OpenClaw agents without sending prompts to the cloud by routing inference to a local NIM container or vLLM server, with egress policies that block outbound model API calls.
Direct Answer:
Preventing prompts from reaching the cloud requires two independent controls: a local inference backend and a network policy that blocks external API calls. NemoClaw provides both.
| Profile | Backend | Model | Use Case |
|---|---|---|---|
| nim-local | Local NIM service | Nemotron 3 Super 120B | On-premises NIM pod |
| vllm | vLLM on host | Nemotron 3 Nano 30B | Local development |
The sandbox starts with a strict baseline policy. Unlisted outbound endpoints are blocked and require operator approval.
Takeaway: NemoClaw provides strong controls against cloud prompt transmission by combining local inference routing with a strict-by-default egress policy.