Fundamentals

What Is NemoClaw?

Last updated: April 1, 2026 · 8 min read

Quick Answer — What is NemoClaw?

NemoClaw is NVIDIA's open-source security reference stack, announced at the GTC conference on March 16, 2026. It wraps the OpenClaw autonomous AI agent inside a sandboxed environment governed by NVIDIA's OpenShell secure runtime — enforcing policy-based privacy guardrails, network egress controls, and per-action intent verification to prevent data exfiltration and shell-level exploits.

The Problem NemoClaw Solves

OpenClaw emerged as one of the most capable autonomous AI agent platforms of 2026. It functions as an operating system for personal AI — capable of running arbitrary shell commands, reading and writing files anywhere on the host system, browsing the web, and taking proactive action across an entire user environment without explicit per-step authorization.

This power introduced an equally extreme attack surface. A single compromised prompt injection could instruct OpenClaw to exfiltrate SSH keys, delete critical system files, install malware, or establish persistent outbound network connections to attacker infrastructure — all while the user believed the agent was performing a routine task.

Enterprises evaluating OpenClaw for deployment in developer environments, CI/CD pipelines, and corporate workflows needed a way to retain the agent's productivity while constraining its blast radius. NVIDIA's answer was NemoClaw.

How NemoClaw Works: OpenShell Runtime

NemoClaw works through two components: an OpenClaw terminal plugin and an OpenShell runtime blueprint. OpenShell sits between the agent and the underlying infrastructure — intercepting every proposed action, verifying it against an operator-defined policy file, and either permitting or denying execution before anything touches the host.

The plugin integrates directly into the OpenClaw terminal subprocess. When the agent proposes an action — file read, shell execution, network request, or model call — OpenShell intercepts it before it executes. It then evaluates the proposed action against a YAML policy file defined by the operator.

This policy file defines what the agent is explicitly allowed to do: which file system paths are accessible, which outbound domains are permitted, which shell commands are allowed, and how personally identifiable information should be handled before any data leaves the local environment. Actions that fall outside the policy are blocked silently or with a configurable error response.

NemoClaw vs Raw OpenClaw

DimensionRaw OpenClawNemoClaw
Primary use casePersonal productivity, local devEnterprise, regulated environments
Shell execution scopeUnrestricted on host OSScoped by operator YAML policy
File system accessFull host read/writeLandlock + seccomp isolation
Outbound networkAny destination, any timeExplicit allowlist per operator
Model call routingDirect to any cloud providerThrough NVIDIA privacy router
PII protectionNone — raw data passed throughStripped before cloud transmission
Agent privilege scopeInherits user session permissionsMinimal-privilege per agent instance
Intent verificationNoneValidated against policy before each action
Audit loggingPersonal / noneStructured audit trail for compliance
Production readinessReference architectureAlpha — not yet production-grade

NVIDIA's OpenClaw Security Reference Stack

NemoClaw is described by NVIDIA as a security reference stack, not a production product. This is significant. It means NemoClaw is designed to show developers how to build secure agentic AI deployments — to be studied, forked, hardened, and extended — rather than deployed directly into production as-is.

The open-source nature of the project means the community can audit every component, identify gaps (several are documented on the security gaps page), and contribute hardening improvements back upstream. NVIDIA has explicitly acknowledged that indirect prompt injection remains the most significant unresolved risk in the current alpha.

For teams evaluating whether to adopt NemoClaw, the honest answer is: use it as a learning framework and as a basis for your own hardened implementation. Do not deploy the unmodified alpha into production systems handling sensitive data.

The Relationship Between NemoClaw and OpenClaw

NemoClaw is an enterprise distribution of OpenClaw. OpenClaw is the brain and the motor; NemoClaw is the cage and the rulebook. You cannot run NemoClaw without OpenClaw — NemoClaw adds security controls on top of the core agent, it does not replace it.

Technologically, NemoClaw's two components — the terminal plugin and the OpenShell blueprint — are designed to be minimally invasive to OpenClaw's core functionality. The goal is that from the agent's perspective, the sandbox is invisible. Actions that fall within the policy succeed without delay. Actions that fall outside the policy fail with configurable errors.

This design philosophy mirrors the principle of least-privilege in traditional systems security: give the agent exactly the permissions it needs for its defined tasks, and nothing more.

Who Announced NemoClaw and When?

NemoClaw was announced by NVIDIA CEO Jensen Huang at the GTC conference on March 16, 2026. It was released as an open-source project under a permissive license, designed to serve as the community benchmark for secure autonomous agent deployment in enterprise environments.

The announcement came at a pivotal moment. OpenClaw had seen viral developer adoption throughout late 2025, with monthly search volumes exceeding 2.3 million queries. But the pace of enterprise adoption was limited by security concerns. NemoClaw was NVIDIA's direct response to those concerns — providing the industry with a credible, GPU-native, open-source security framework built by the same organization that provides the inference hardware underpinning most frontier AI deployments.

Next up

NemoClaw Architecture →

Setup

Install NemoClaw →

Security

Known Security Gaps →