Nvidia’s NemoClaw is OpenClaw with guardrails
Nvidia's NemoClaw combines the OpenClaw agent platform with components of its Agent Toolkit to add privacy and security controls.
by Frederic Lardinois · The New StackAt its annual GTC conference, Nvidia on Monday announced the Nvidia Agent Toolkit, which brings together open models, runtimes, open skills, and blueprints to build long-running, secure, and performant autonomous agents.
Essentially, this is the next generation of what the company previously called the Nvidia NeMo Agent Toolkit. While Agent Toolkit is interesting in itself, it’s NemoClaw that will likely steal most of the attention.
NemoClaw was rumored to be somewhat of an enterprise competitor to OpenClaw, the autonomous AI agent framework that became the fastest-growing open-source project in history. In reality, it is a project that integrates OpenClaw into Nvidia’s open agentic stack, with the Agent Toolkit providing some of the missing parts, especially for security and guardrails for OpenClaw agents.
In its announcement, Nvidia calls it “the Nvidia NemoClaw stack for the OpenClaw agent platform.”
Nvidia says it developed NemoClaw in collaboration with OpenClaw founder Peter Steinberger, who remains the maintainer of OpenClaw even after he joined OpenAI earlier this year.
“Every company now needs to have an OpenClaw strategy,” Nvidia CEO Jensen Huang said in his keynote. To him, OpenClaw — and claws in general — are going to be as important as Linux, Kubernetes, HTML, and other fundamental tools.
NemoClaw = OpenClaw for the enterprise with privacy and security guardrails
The idea here is for NemoClaw to become something akin to an enterprise-grade distribution of OpenClaw. It installs with a single command and is designed to provide a secure version of OpenClaw.
NemoClaw can use any coding agent. While OpenClaw handles the runtime, memory, and skills, NemoClaw adds new and existing open source models, tools, and frameworks from Nvidia.
It can use, for example, Nvidia’s own Nemotron models (or any other model running locally or in the cloud), the company’s Dynamo inference engine, and a new open-source security runtime called OpenShell that is at the core of the Agent Toolkit.
OpenShield and the Agent Toolkit
“Nvidia OpenShell is a new open source safety and security runtime for agents,” Kari Briski, Nvidia’s VP of generative AI software. said in a press conference ahead of today’s announcement. “OpenShell provides the missing infrastructure layer beneath claws to give them the access they need to be productive, while enforcing policy-based security network and privacy guardrails.”
This security layer is at the core of the announcement. As claws gain access to corporate tools and data, OpenShell is designed to be the policy-enforcement layer that keeps them within bounds by combining security, network, and privacy guardrails.
Nvidia, in its announcement, argues, “This provides the missing infrastructure layer beneath claws to give them the access they need to be productive, while enforcing policy-based security, network, and privacy guardrails.”
It’s no secret that OpenClaw had its share of security issues, especially in its earliest iterations. Many of these have been fixed by now, but the risk of giving an autonomous agent access to private or enterprise data and tools isn’t something a bug fix can correct. That’s where something like OpenShield comes in, which will keep the agent within its guardrails.
“NemoClaw installs OpenClaw with Nemotron models and the Nvidia OpenShield runtime in a single command,” Briski explained. “This provides a foundation for agents to develop and learn new skills to complete tasks according to defined privacy and security guardrails. It adds security and privacy to run personal, always-on AI assistance anywhere.”
Nvidia is working with Cisco, CrowdStrike, Google, Microsoft Security, and TrandAI to bring OpenShell compatibility to their security tools
NemoClaw can run in the cloud, as well as locally on RTX PCs and Nvidia’s own desktop supercomputers like DGX Spark and DGX Station. Some enterprising developers will likely find ways to run this on the Mac minis they bought for OpenClaw, too.
Indeed, with this launch, Nvidia is now positioning those systems as dedicated development platforms for building claws. The DGX Spark, as well as similar systems from its partner, have been on the market for a while. The DGX Station, the higher-end option for running frontier-class models locally, opens for orders on Monday.
Nvidia and open source
During a press conference ahead of today’s announcement, Briski framed this move as part of Nvidia’s ongoing engagement with the open source community.
“Nvidia NemoClaw, which you’ve all already seen in the news, is Nvidia’s contribution to the open claw community to help take the incredible OpenClaw phenomenon to the next level. Just like we’ve done for Pytorch, Kubernetes, OpenGL, and more,“ she says.