NVIDIA/NemoClaw

Run agents like Hermes, LangChain Deep Agents, and OpenClaw more securely inside NVIDIA OpenShell with managed inference

NVIDIA NemoClaw – Reference Stack for Sandboxed AI Agents

What it is – An open‑source toolkit that lets you run “always‑on” AI agents (e.g., OpenClaw, Hermes, LangChain Deep Agents) inside the NVIDIA OpenShell sandbox environment. It bundles a hardened sandbox blueprint, network‑policy controls, routed inference configuration, and a single‑command‑line interface for installing, configuring, and managing the agents safely.

Why it matters – Running autonomous agents continuously poses security risks (uncontrolled network access, credential leakage, runaway processes). NemoClaw addresses these concerns by:

  • Providing a pre‑hardened container sandbox with capability drops and resource limits.
  • Enforcing network egress policies that require operator approval for outbound connections.
  • Routing inference calls through vetted providers (local or cloud) with validation.
  • Offering a unified CLI that handles lifecycle (create, start, stop, repair) and policy updates.

Supported agents

  • OpenClaw – default open‑source agent.
  • Hermes – another popular conversational agent.
  • LangChain Deep Agents (Code) – agents built with LangChain’s deep‑agent framework.

Key components

Component Role
Plugin/Blueprint Defines the sandbox image, capabilities, and security hardening.
Network Policy Engine Baseline egress rules + operator‑approval flow; customizable presets.
Inference Router Selects and validates an inference provider (e.g., local model, NVIDIA AI services).
NemoClaw CLI nemo command for install, quick‑start, sandbox management, and policy tweaks.
Documentation site Detailed architecture, security best practices, and step‑by‑step guides.

Typical workflow

  1. Install – Run the interactive installer (curl … | bash) after checking prerequisites.
  2. Choose an agent – Set NEMOCLAW_AGENT=hermes (or leave default) before install, or use the nemohermes alias after.
  3. Start a sandbox – nemo start creates a hardened container and launches the selected agent.
  4. Interact – Communicate with the agent via your preferred UI (terminal, IDE, or coding assistant) while the sandbox enforces network and credential policies.
  5. Manage – Use nemo stop, nemo repair, or nemo policy edit to control lifecycle and security settings.

Getting started – The README points to a “starter prompt” you can feed to a coding assistant (e.g., Cursor, Claude Code) that will walk you through installation. Quick‑start guides for each supported agent are linked directly from the README.

Community & support – Alpha‑stage project with:

  • GitHub Discussions and Issues for questions and bug reports.
  • A Discord server for real‑time help.
  • Private channels for security vulnerability disclosure (see SECURITY.md).

License – Apache 2.0 (permissive, allowing commercial use and modification).

Bottom line – NemoClaw is a practical, security‑focused stack that lets developers and enterprises run persistent AI agents inside a controlled sandbox, reducing the attack surface while keeping the agents functional and easy to manage.