NVIDIA NemoClaw: OpenClaw Plugin for Sandboxed Autonomous Agent Deployment

2026-03-18T17:21:39.000Z·2 min read
NVIDIA releases NemoClaw, an alpha-stage OpenClaw plugin that combines NVIDIA OpenShell runtime with Nemotron models for secure, sandboxed deployment of autonomous AI agents with declarative policy enforcement across network, filesystem, process, and inference layers.

NVIDIA has released NemoClaw (alpha), an open-source plugin for OpenClaw that simplifies running always-on AI assistants safely by combining the NVIDIA OpenShell runtime with Nemotron models in a fully sandboxed environment.

What It Does

NemoClaw orchestrates a complete stack for autonomous agents:

ComponentRole
PluginTypeScript CLI for launch, connect, status, logs
BlueprintVersioned Python artifact for sandbox creation and policy
SandboxIsolated OpenShell container with policy-enforced egress and filesystem
InferenceNVIDIA cloud model calls routed through OpenShell gateway

Security Layers

The sandbox enforces four protection layers:

When the agent tries to reach an unlisted host, OpenShell blocks the request and surfaces it in the TUI for operator approval.

Requirements

ResourceMinimumRecommended
CPU4 vCPU4+ vCPU
RAM8 GB16 GB
Disk20 GB free40 GB free
OSUbuntu 22.04+
Node.js20+
DockerInstalled and running

Quick Start

curl -fsSL https://nvidia.com/nemoclaw.sh | bash
nemoclaw my-assistant connect

The onboard wizard handles sandbox creation, inference configuration, and security policy setup.

Inference Model

Uses nvidia/nemotron-3-super-120b-a12b via NVIDIA Cloud API. Inference requests never leave the sandbox directly — OpenShell intercepts and routes all calls.

Alpha Status

NemoClaw is explicitly labeled as early-stage alpha. Interfaces and APIs may change. NVIDIA is gathering feedback for iteration toward production readiness.

Source: GitHub - NVIDIA/NemoClaw | HN Discussion

↗ Original source
← Previous: Google Engineers Launch 'Sashiko' for Agentic AI Code Review of the Linux KernelNext: Snowflake Cortex AI Sandboxed Escape: Prompt Injection Bypasses Human-in-the-Loop →
Comments0