NVIDIA/NemoClaw ? reverse-engineered prompt

I want to build a command-line tool that helps me run open-source AI agents, like OpenClaw, more safely. The main idea is to create a secure wrapper around the agent using a runtime like NVIDIA's OpenShell. This should put the agent in a sandbox, controlling its access to my file system and network to prevent it from doing anything unexpected.

The tool needs a guided onboarding process. A command like nemoclaw onboard should walk me through creating a new sandboxed agent, connecting it to a powerful inference model from a provider like NVIDIA, and applying some sensible default security policies.

Once an agent is running, I'll need simple commands to connect to its shell, check its status, and see its logs. Let's also include a basic terminal UI for chatting with the agent once I'm connected. The whole thing should be easy to install with a single curl script and should work across Linux, macOS, and Windows with WSL2, assuming I've already got Docker installed and running.