A Warden for AI Agents That Never Sleep

What it does

NemoClaw is an open-source reference stack that runs persistent AI agents inside NVIDIA OpenShell sandboxes. It bundles guided onboarding, sandbox lifecycle management, routed inference, and network policy into a single CLI. Out of the box it supports OpenClaw and Hermes agents.

The interesting bit

Instead of treating security as a bolt-on, the project bakes capability drops, process limits, and egress controls into a hardened blueprint. It is essentially infrastructure-level guardrails for operators who do not trust their agents with unchecked network access.

Key highlights

• Sandboxed execution inside NVIDIA OpenShell with hardened container blueprints
• Single CLI manages onboarding, inference routing, and agent lifecycle
• Supports OpenClaw (default) and Hermes via the nemohermes alias
• Network policies include operator approval flows and baseline egress rules
• Apache 2.0 licensed

Caveats

• Substantive architecture, hardening, and setup details live almost entirely on external NVIDIA documentation pages, not in the repository
• The disclaimer explicitly warns that the software automatically retrieves external materials and users bear full responsibility for their security, integrity, and licensing
• Tightly coupled to the NVIDIA OpenShell ecosystem

Verdict

Worth evaluating if you run long-lived AI agents and want a prescriptive, locked-down posture. Look elsewhere if you need a standalone agent framework or prefer to avoid ecosystem lock-in.