Jensen Huang: every company needs an OpenClaw strategy
At GTC 2026, NVIDIA CEO Jensen Huang spent three hours on stage with a clear message: AI agents are the future of software. His exact words: "OpenClaw is the operating system for personal AI."
The headline announcement: NVIDIA NemoClaw. An open-source security stack built on top of OpenClaw, developed in collaboration with OpenClaw creator Peter Steinberger.
NVIDIA expects 10 million digital workers operating alongside humans. NemoClaw is how they plan to make that safe.
What NemoClaw actually is
NemoClaw is not a replacement for OpenClaw. It wraps OpenClaw inside a hardened, policy-enforced container environment using NVIDIA OpenShell. Think of it as enterprise armor for your AI agent.
The architecture has four components:
| Component | What it does |
|---|---|
| Plugin | TypeScript CLI for launching, connecting, and monitoring sandboxes |
| Blueprint | Python orchestration layer managing sandbox creation and policy setup |
| Sandbox | Containerized OpenShell environment running OpenClaw with enforced policies |
| Inference | Routes AI model calls through controlled backends with policy checks |
Installation is a single command:
curl -fsSL https://nvidia.com/nemoclaw.sh | bash
The installer handles Node.js, Docker, and OpenShell setup automatically.
Security features that matter for businesses
NemoClaw addresses the biggest concern companies have about autonomous AI agents: what happens when an agent goes off-script?
Network policies - Block unauthorized outbound connections. Your agent cannot call APIs you have not approved.
Filesystem restrictions - Agents can only access /sandbox and /tmp. No reading your entire server.
Process controls - No privilege escalation. The agent cannot sudo, install packages, or modify system files outside its sandbox.
Inference interception - All model calls route through controlled backends. You decide which models the agent can use and what prompts are allowed.
Hot-reloadable policies - Change security rules at runtime without restarting your agent.
Local AI models with Nemotron
NemoClaw supports NVIDIA Nemotron models for fully local inference. Your data never leaves your hardware.
This is a big deal for industries with strict data requirements: healthcare, legal, finance, government. The agent runs on your GeForce RTX PC, RTX PRO workstation, or DGX hardware. All processing stays local.
When local compute is not enough, NemoClaw includes a privacy router that connects to cloud models within your defined safety parameters. You set the rules about what data can leave your network.
What this means for OpenClaw users
If you already run OpenClaw - like we do at AI Agent - NemoClaw is an upgrade path, not a migration. Your existing OpenClaw setup continues to work. NemoClaw adds a security layer around it.
For businesses evaluating AI agents, the equation just changed. The "but is it secure enough?" objection has an answer now: NVIDIA-backed, open-source, policy-driven security with full audit trails.
The source code is available on GitHub: github.com/NVIDIA/NemoClaw (Apache 2.0 license).
AI Agent and NemoClaw
At AI Agent B.V. we have been working with OpenClaw since it launched. We build, deploy, and manage OpenClaw agents for businesses across the Netherlands and Europe.
With NemoClaw, we are adding enterprise security capabilities to our platform. Customers who need stricter data controls - local inference, network policies, sandboxed execution - can now get that without leaving the OpenClaw ecosystem.
We are OpenClaw and NemoClaw specialists. Whether you want a personal assistant, a customer service agent, or a full multi-agent setup with enterprise security, we can help.
Getting started
Self-service: Launch your own OpenClaw agent at aiagent.nl/openclaw. Dedicated EU server, GDPR compliant, live in minutes.
Custom implementation: Need NemoClaw integration, local models, or enterprise security? Contact us for a custom setup.
Advisory: Not sure where to start? Book a free 30-minute call at aiagent.nl/gesprek.
