Announced at NVIDIA GTC · March 16, 2026

NemoClaw: NVIDIA's Security Stack for OpenClaw Autonomous Agents

Name: NemoClaw
Author: NVIDIA

What is NemoClaw?

NemoClaw is NVIDIA's open-source security reference stack that sandboxes the OpenClaw autonomous AI agent inside a controlled environment governed by the OpenShell secure runtime. It enforces policy-based privacy, network egress controls, and intent verification — preventing shell-level exploits and data exfiltration attacks.

Install NemoClaw Explore Architecture

2.3M+

Monthly OpenClaw searches

Security isolation layers

Mar 2026

GTC announcement date

Open Source

MIT licensed reference stack

What Is NemoClaw?

NemoClaw is an open-source security reference stack from NVIDIA that wraps the OpenClaw autonomous agent inside an isolated, policy-governed sandbox. It uses NVIDIA's OpenShell secure runtime to control what the agent can access, execute, and communicate — without removing its core capabilities.

OpenClaw gained mass traction as an autonomous AI agent capable of operating directly on a user's local hardware or cloud infrastructure. Because it can run shell commands, interact with files, and browse the web, developers recognized extreme security risks for enterprise environments.

NVIDIA answered this by developing NemoClaw — wrapping the OpenClaw agent inside a controlled sandbox governed by the OpenShell secure runtime. OpenShell sits between the agent and the underlying infrastructure, controlling exactly what the agent can access, execute, and where model calls are sent.

NemoClaw works as a plugin that integrates with the OpenClaw terminal and a blueprint that orchestrates all resources through OpenShell. Every action the agent proposes is verified against an operator-defined YAML policy file before execution is permitted.

Full technical explanation of NemoClaw →

NemoClaw vs OpenClaw

Capability	OpenClaw	NemoClaw
Shell command execution	Unrestricted	Policy-gated
File system access	Full host access	Landlock sandboxed
Network outbound	Unrestricted	Allowlist only
PII handling	No stripping	Privacy router active
Intent verification	None	Per-action policy check
Production ready	Reference only	Alpha — not yet

Full OpenClaw vs NemoClaw deep-dive →

NemoClaw Security Architecture

NemoClaw implements five distinct security layers between the OpenClaw agent and the underlying infrastructure. Each layer addresses a specific attack surface — from raw file system access to cloud model data leakage. All layers operate under the OpenShell runtime policy enforcement engine.

Security Layer	Mechanism of Protection	Vulnerability Addressed
Sandboxed Execution	Isolated containers with Landlock + seccomp filters	Prevents direct access to host file system or adjacent processes
Network Egress Control	All outbound connections blocked unless explicitly approved by operator policy	Prevents data exfiltration attacks where agents phone home
Minimal-Privilege Access	Least-privilege principles enforced per individual agent instance	One compromised agent cannot inherit permissions of another
Privacy Router	Strips personally identifiable information before cloud model calls	Prevents leakage of sensitive user data to third-party cloud models
Intent Verification	Validates agent's proposed actions against operator-defined policy	Forces injection attacks to produce an in-policy action to succeed

Deep-dive architecture documentation →Known security gaps in alpha →

Install NemoClaw

NemoClaw supports Linux natively via a single installer script. macOS requires Docker Desktop or Colima. Windows users must first install WSL2, then use the Docker Desktop backend. A remote VM path is available for teams requiring GPU inference at cloud scale.

The fastest path is Linux with Docker running and at least 8 GB of RAM:

# Linux — fastest install path
curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash

# Verify installation
nemoclaw --version

# Connect your first assistant
nemoclaw my-assistant connect

# Launch the OpenClaw terminal UI
openclaw tui

Complete multi-OS install guide →

Supported Installation Paths

Operating System	Prerequisites	Status
Linux	Docker + 8 GB RAM	Supported
macOS (Apple Silicon)	Docker Desktop / Colima	Supported
Windows	WSL2 + Docker Desktop	Supported
Remote VM (Ubuntu)	24 GB VRAM	GPU Required
GitHub Codespaces	Free tier	Cloud Deploy

Windows WSL2 step-by-step guide →

NemoClaw Documentation Hub

Everything you need to understand, deploy, secure, and extend NemoClaw. Choose a topic.

Fundamentals

What Is NemoClaw?

Full technical explanation of how NemoClaw sandboxes OpenClaw with NVIDIA OpenShell, covering its history and design goals.

Architecture & OpenShell Runtime

Deep dive into the five-layer security stack, the OpenShell runtime, and how NemoClaw integrates as an OpenClaw plugin.

Install NemoClaw

Complete installation guide for Linux, macOS, Windows, remote VMs, and GitHub Codespaces. Terminal commands included.

NemoClaw on Windows (WSL2)

Step-by-step Windows installation guide. Covers WSL2 setup, Docker Desktop backend configuration, and running your first agent.

Security Gaps & Hardening

Critical analysis of alpha vulnerabilities: indirect prompt injection, policy drift, multi-turn erosion, and supply chain risks.

OpenClaw vs NemoClaw

Detailed comparison of OpenClaw's raw capabilities versus NemoClaw's security-first constraints. When to use each.

NemoClaw Alternatives

How NemoClaw competes with Claude Cowork, ChatGPT Agent, Manus, Kuse Cowork, and Eigent AI in 2026's agent landscape.

Download NemoClaw

Official install scripts, GitHub repository links, Docker images, and quick-start commands to get NemoClaw running now.

Known Security Gaps in NemoClaw Alpha

NemoClaw is alpha software and is not production-ready. Its intent verification layer focuses on proposed actions rather than content returned by external tools, creating an attack surface for indirect prompt injection, policy drift, multi-turn erosion, and supply chain threats via community skills.

Security Gap	Technical Mechanism	Mitigation Strategy
Indirect Prompt Injection	Malicious data from trusted sources bypasses intent verification by entering the agent reasoning chain	Sanitize all context windows before execution; treat external tool outputs as untrusted input
Policy Definition Quality	Overly permissive YAML operator policies allow attacks to pass through intent verification	Use hardened baseline policy templates; enforce policy review before any deployment
Multi-Turn Erosion	Intent verification validates per-action only — not cumulative behavioral patterns across sessions	Integrate continuous state audits; monitor behavioral drift across multi-turn agent sessions
Supply Chain Attacks	Vulnerable community-built third-party skills on ClawHub can compromise the agent environment	Use only vetted skills; pin skill versions; review source code before enabling any skill

Full security hardening and red team guide →

Frequently Asked Questions About NemoClaw

What is NemoClaw?

NemoClaw is an open-source security reference stack released by NVIDIA 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, providing policy-based privacy guardrails and network egress controls.

What is the difference between NemoClaw and OpenClaw?

OpenClaw is the core autonomous agent platform — it operates on a user's local hardware, runs shell commands, interacts with files, and browses the web. NemoClaw is the security layer built on top of OpenClaw. It adds Landlock sandboxing, intent verification, and a privacy router to prevent data exfiltration and unauthorized access.

Is NemoClaw production-ready?

No. As of April 2026, NemoClaw is alpha software. Known gaps include susceptibility to indirect prompt injection attacks and reliance on well-crafted operator policy YAML files. It is intended as a reference stack, not a hardened production deployment.

How do I install NemoClaw on Linux?

Ensure Docker engine is running and you have at least 8 GB of RAM, then run: curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash. For Windows, install WSL2 first via PowerShell as Administrator.

What is openclaw?

OpenClaw is an autonomous AI agent platform that acts as an operating system for personal AI. It can run shell commands, interact with files, browse the web, and take proactive actions across a user's machine. Because of these broad capabilities, NVIDIA developed NemoClaw as a secure runtime wrapper around it.

Ready to run NemoClaw?

Follow the complete installation guide for your operating system. Linux, macOS, Windows WSL2, remote VMs, and GitHub Codespaces all supported.

Installation Guide Download Scripts