NVIDIA’s NemoClaw Gives Personal AI Agents the Safety Companies Need

OpenClaw took the tech industry by surprise earlier this year when an Austrian engineer created the first version in roughly an hour. This small project swept through the community like wildfire as the most open-source endeavor on record, allowing anyone to set up a personal AI agent to operate directly on their own PC and accomplish tasks like organizing files or pounding out code without sending any data off to who knows where. At GTC, NVIDIA introduced NemoClaw, a software add-on for OpenClaw that can be installed with a single command. This new layer includes the security features and privacy controls that transform these agents from fun little experiments to useful business solutions.

NemoClaw is simple to integrate into your existing setup since it introduces OpenShell, a runtime that isolates each agent in its own small bubble. Then you can create rules in plain text files that specify which folders the agent may browse, which networks it can connect to, and which external services it can access. Everything else, and we mean everything, is off limits, and every step they make leaves a clear paper trail for you to follow.

NVIDIA is marrying this with some of their own open models known as Nemotron, which run locally on whatever hardware is available. You’re talking everything from RTX-powered laptops to entire workstations and dedicated AI systems. When you need a little more horsepower, you can utilize a privacy router to connect to more powerful cloud models while keeping all of your sensitive information in-house. The end result is a framework that allows your agents to work in a mix of local and remote resources while maintaining tight boundaries.

Companies are already putting these components to use in real-world situations. Cisco runs agents that detect security flaws, verify databases, map affected devices, and create a thorough remediation plan, all of which must be checked against the rules in real time. Box uses the same framework to handle invoices and contracts, using abilities that work well with existing access levels. NVIDIA has also partnered with Salesforce, CrowdStrike, and a few more large brands to achieve the same level of control across all of their technologies. OpenClaw agents run 24 hours a day, seven days a week on personal computers, professional workstations, and servers without interfering with anything else. If you have dedicated hardware, you can keep them up and running for hours, if not days, while they work.

It is currently available as an early preview through official sources, such as on Github. Users can begin testing right away, although NVIDIA says some rough edges must be ironed out before they are formally put into production. Developers have complete access to the toolkit, which includes sample models, runtimes, and guidance for creating their own agents. Jensen Huang referred to OpenClaw as the operating system for personal AI, indicating a trend toward software that can be instructed to do things. Peter Steinberger, the original creator of OpenClaw, sees this combination as a method for users to design and run their own secure assistants.