NVIDIA NemoClaw is an open source reference stack written in TypeScript that enables secure execution of AI agents within NVIDIA OpenShell sandboxes. The project provides a comprehensive solution for running always-on AI agents like OpenClaw, Hermes, and LangChain Deep Agents Code with enhanced safety controls and managed inference capabilities. The repository serves as a hardened blueprint that combines guided onboarding, routed inference, network policy enforcement, and lifecycle management through a unified command-line interface.
The core functionality centers on sandboxing and security. NemoClaw delivers sandbox hardening through container security measures, capability drops, and process limits. It implements network policies with baseline rules, operator approval flows, and egress control mechanisms. The system supports customizable network policies with both static and dynamic changes, along with presets for different security postures. Documentation covers security best practices, a controls reference, and risk frameworks to help operators configure sandbox security appropriately for their threat models.
Inference management is a key architectural component. NemoClaw supports multiple inference providers and includes routed inference configuration to direct agent requests appropriately. The system validates inference behavior across supported provider paths and maintains documentation on inference options and provider selection. This allows operators to choose between local inference and routed inference depending on their infrastructure and security requirements.
The project emphasizes accessibility through multiple onboarding paths. Users can start with an interactive terminal installer, use a starter prompt designed for coding agents like Cursor and Claude Code, or follow quickstart guides specific to each supported agent. The documentation is extensive, covering prerequisites, architecture details, CLI commands, troubleshooting, and ecosystem integration with OpenClaw and OpenShell. The repository includes a contributor path separate from the end-user installer, with development setup scripts and coding standards documented in CONTRIBUTING.md.
Community engagement occurs through GitHub Discussions, Discord, and GitHub Issues. The project explicitly identifies itself as an alpha project where maintainers review contributions on a best-effort basis.
Current priorities focus on improving install and onboarding reliability across tested platforms, strengthening sandbox hardening and credential handling, validating inference behavior for supported providers, and keeping documentation aligned with supported workflows. Security vulnerabilities must be reported through private channels via the NVIDIA Vulnerability Disclosure Program, email to [email protected], or GitHub's private vulnerability reporting rather than public issues. The project is licensed under Apache 2.0 and includes a notice that the software may retrieve external materials governed by separate terms and conditions.
