Nvidia has launched NemoClaw, a software stack for the OpenClaw agent platform that installs Nvidia Nemotron models and the new Nvidia OpenShell runtime in a single command.
Announced at its AI conference in San Jose, Nvidia said NemoClaw adds privacy and security controls for self-evolving, autonomous agents running from cloud and on-premises environments to RTX PCs, DGX Station and DGX Spark systems.
NemoClaw uses Nvidia Agent Toolkit software to optimize OpenClaw, installs OpenShell to provide open models and an isolated sandbox, and adds policy-based security, network and privacy guardrails.
Agents can use open models running locally on dedicated systems and, through a privacy router, frontier models running in the cloud.
“With Nvidia and the broader ecosystem, we’re building the claws and guardrails that let anyone create powerful, secure AI assistants,” Peter Steinberger, creator of OpenClaw said in the announcement.
The NeMo and AI-Q layer underneath NemoClaw
NeMo provides microservices and toolkits for data processing, model fine-tuning and evaluation, reinforcement learning, policy enforcement, and system observability.
Its AI-Q blueprint separately describes AI-Q as an open reference example for building AI agents that developers can own, inspect and control, with benchmarks and evaluation harnesses to measure quality and improve over time.
How Microsoft and OpenAI are approaching the same problem
The timing also fits a broader shift in the agent market. Microsoft said last year that Copilot Studio was adding managed security enhancements centered on proactive governance, secure-by-default controls and comprehensive visibility, including controls over connectors, sharing scopes and authentication.
Microsoft later added advanced real-time protection during agent runtime, describing it as a way for organizations to connect monitoring systems and evaluate and control agent behavior in real time as agents take on more autonomous tasks.
OpenAI has been moving in a similar direction for coding agents. In February, the company introduced the Codex app as a way to manage multiple agents, run work in parallel and supervise long-running tasks, presenting it as a control surface for agent teams rather than just a single-model interface.
NemoClaw is being positioned around local and hybrid execution, isolated sandboxes and a privacy router and policy-based guardrails for open agents that can stay on dedicated hardware around the clock.
The guardrails layer and what Nvidia says it fills
Nvidia says NeMo Guardrails lets developers define rails that guide and constrain LLM and agent behavior, and the library includes content safety, topic control and execution guardrails.
Nvidia says those protections are meant to provide the “missing infrastructure layer” beneath claws, giving agents the access they need to be productive while enforcing privacy and network policies.