<p>Nvidia announced NemoClaw at GTC 2026, an enterprise-grade distribution of the open-source OpenClaw AI agent framework. The project was developed in collaboration with OpenClaw creator Peter Steinberger, who described the work as “so much fun cooking” with the Nvidia team.</p>
<h2>What it is</h2>
<p>OpenClaw is an open platform for connecting AI agents to applications and local tools. It’s gained traction for enabling automation, but that same capability raised real security concerns: persistent shell access, live credentials, and self-modifying tooling create a meaningful attack surface when agents run unattended.</p>
<p>NemoClaw wraps OpenClaw with Nvidia’s Agent Toolkit and adds OpenShell, an open-source safety runtime that sits between the agent and the underlying infrastructure. The key design decision is that OpenShell runs <em>out-of-process</em> — meaning the agent itself cannot override the constraints, even if it’s compromised or misbehaving.</p>
<h2>How OpenShell works</h2>
<p>OpenShell has three main components:</p>
<ul>
<li>An isolated sandbox for skill development, so agents can’t reach outside their designated environment</li>
<li>A policy engine that evaluates actions at the binary, network, and filesystem level before they execute</li>
<li>A privacy router that keeps sensitive data on-device and only routes to external frontier models when policy explicitly permits it</li>
</ul>
<p>Agents like Claude Code, Codex, and Cursor can run inside OpenShell unmodified. NemoClaw installs with a single command.</p>
<h2>Hardware and model support</h2>
<p>NemoClaw supports Nvidia’s Nemotron models and the Dynamo inference engine, but it’s hardware agnostic and works with any open AI model. It runs in the cloud or locally on GeForce RTX PCs, RTX Pro workstations, DGX Station, and DGX Spark systems.</p>
<p>Nvidia is working with Cisco, CrowdStrike, Google, Microsoft Security, and TrandAI on OpenShield compatibility.</p>
<h2>Where it fits</h2>
<p>NemoClaw is currently in early alpha. It puts Nvidia in the growing market for enterprise AI agent governance platforms, alongside OpenAI Frontier and others, targeting companies that want to deploy autonomous agents but are understandably nervous about what those agents might do with access to corporate tools and sensitive data.</p>


Collections

Nvidia announced NemoClaw at GTC 2026, an enterprise distribution of the open-source OpenClaw AI agent framework. It adds OpenShell, a safety runtime that runs out-of-process between AI agents and underlying infrastructure, preventing agents from overriding security constraints even if compromised. OpenShell includes an isolated sandbox, a policy engine evaluating actions at binary/network/filesystem level, and a privacy router keeping sensitive data on-device. Agents like Claude Code, Codex, and Cursor run inside it unmodified. NemoClaw supports Nvidia hardware and models but is hardware-agnostic, and Nvidia is partnering with Cisco, CrowdStrike, Google, and Microsoft Security on OpenShield compatibility. Currently in early alpha, it targets enterprises wanting to deploy autonomous agents without exposing corporate tools and sensitive data.

NVIDIA's NemoClaw adds security guardrails to OpenClaw AI agents