NVIDIA/NemoClaw

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

NVIDIA NemoClaw – What It Is

NemoClaw is an open‑source reference stack that lets you run AI agents (like OpenClaw, Hermes, or LangChain Deep Agents) inside NVIDIA OpenShell sandboxes. It adds a thin management layer on top of OpenShell that handles:

  • Guided onboarding and a CLI for creating, starting, and stopping agent sandboxes
  • Managed inference routing to supported model providers
  • Network‑policy enforcement and container hardening to keep the sandbox safe
  • State persistence backed by OpenShell so an agent can keep its memory across runs
  • Easy switching between the three supported agents via CLI aliases

The project is positioned as a safety‑focused “wrapper” for agents that need to execute code, access the internet, or use local resources, making them easier to run on DGX systems or WSL.

Key Features (as described in the README)

Feature What It Does
Supported agents OpenClaw (default), Hermes, LangChain Deep Agents Code.
Installer & quick‑start An interactive terminal installer that can run in an “express” mode with preset settings for DGX or WSL.
CLI & aliases A nemoclaw command‑line tool plus agent‑specific shortcuts for common lifecycle actions (create, start, stop, delete, etc.).
Inference provider selection Built‑in wizard to pick a local or routed inference service and configure it for the sandbox.
Network policies Baseline egress rules, operator‑approval flow, and the ability to customise static or dynamic policies.
Sandbox hardening Container security defaults (capability drops, process limits, etc.) and a “hardening” checklist.
State persistence OpenShell‑backed storage so an agent’s memory or files survive sandbox restarts.
Managed skins For Hermes, a lightweight nemoclaw-light skin is installed on demand and removed when not needed.
Documentation & community Full online docs, Discord, GitHub Discussions, and a community‑examples repo.
Security‑first process Private vulnerability reporting channels and a clear disclaimer about external material handling.

Who Might Use It

  • ML engineers or data scientists who want to experiment with autonomous agents but need a controlled environment that limits network access and container privileges.
  • Enterprises deploying coding assistants (e.g., Claude Code, Copilot) on‑premise and requiring sandboxed execution on NVIDIA hardware.
  • Researchers building new agent‑to‑agent workflows who want a ready‑made integration with OpenShell and a way to swap between OpenClaw, Hermes, or LangChain agents without rewriting plumbing code.
  • DevOps / security teams that need a reproducible, auditable way to run agents while enforcing network‑policy and hardening standards.

Getting Started (from the README)

  1. Check prerequisites – see the “Prerequisites” doc page for supported hardware (DGX, WSL, etc.) and software requirements.
  2. Run the installer – clone the repo and execute the provided installer script. On a supported host you can simply press Enter at the prompt to accept the recommended settings.
  3. Choose an agent – the installer defaults to OpenClaw; you can type n during install to pick Hermes or LangChain Deep Agents Code, name the sandbox, and select an inference provider.
  4. Follow the quick‑start guide – each agent has its own “Quickstart” doc page with the exact CLI commands to spin up a sandbox and start chatting with the agent.
  5. Manage the sandbox – use nemoclaw <command> (e.g., nemoclaw start <sandbox>, nemoclaw stop <sandbox>, nemoclaw logs <sandbox>) to control the lifecycle.

Development & Contribution

  • Run ./scripts/dev-setup.sh to set up a source checkout without launching a sandbox.
  • Add --expose-cli if you need the CLI visible on the host during development, or --with-runtime to test sandbox‑related changes.
  • Contributions are welcomed via pull requests; the repo includes a CONTRIBUTING.md with coding standards and PR workflow.

Where to Find More

License

Apache 2.0 (per the LICENSE file).

Related

  • Project
    nanocoai/nanoclaw

    NanoClaw is a lightweight, open‑source framework for running AI agents (Claude, OpenAI, Ollama, etc.) in isolated Docker containers. It provides secure per‑agent sandboxing, multi‑channel messaging (Slack, Discord, WhatsApp, etc.), a credential‑gateway for secret handling, scheduled tasks, and a code‑assistant‑driven setup/debug flow. The core codebase is intentionally small, encouraging users to customize behavior by editing code (or asking Claude Code to do it) rather than managing sprawling config files.

  • Project
    onecli/onecli

    OneCLI is an open-source platform that allows teams to deploy and manage secure, sandboxed AI agents for every employee while centralizing credential management and security policies.

  • Project
    NVIDIA/OpenShell

    A private, policy-enforced runtime that lets AI agents safely read files, call APIs, and use credentials by confining each agent to an isolated sandbox and formally verifying any policy change before it goes live.

  • Project
    opensandbox-group/OpenSandbox

    OpenSandbox provides isolated, secure environments on your own infrastructure for AI agents to execute code, run commands, and operate browsers or desktops.

  • Project
    swarmclawai/swarmclaw

    A self-hosted AI agent runtime and multi-agent framework for creating autonomous agent swarms with durable memory, scheduling, and multi-provider LLM support.