NVIDIA/NemoClaw
OfficialRun OpenClaw more securely inside NVIDIA OpenShell with managed inference
NVIDIA NemoClaw is NVIDIA's open-source reference stack for running OpenClaw always-on AI assistants safely inside NVIDIA OpenShell. Layers guided onboarding, a hardened blueprint, state management, and routed inference on top of the OpenShell sandbox runtime.
Our Review
NVIDIA NemoClaw is NVIDIA's open-source reference stack for running OpenClaw always-on AI assistants safely inside NVIDIA OpenShell -- 20,500 stars and 2,700 forks in two months as of May 2026. Released March 2026 by NVIDIA under Apache 2.0 and currently in alpha preview, it installs the OpenShell runtime (part of NVIDIA Agent Toolkit) and layers guided onboarding, a hardened blueprint, state management, and routed inference on top.
What NVIDIA NemoClaw provides:
- •Hardened OpenShell sandbox containerized execution environment for running OpenClaw agents with additional security boundaries.
- •Guided onboarding step-by-step installer that handles Docker, k3s, OpenShell gateway, and the OpenClaw blueprint in one flow.
- •Routed inference managed inference routing so your OpenClaw agents call the right NVIDIA-hosted or local model per task.
- •State management persistent state for long-running always-on agents that need memory across sessions.
- •Channel messaging OpenShell-managed messaging channels for sending and receiving agent communication.
- •Reference blueprint complete declarative blueprint for the standard OpenClaw stack you can fork and customize.
- •Layered protection multiple defense layers between the autonomous agent and the host system, recommended over running OpenClaw bare.
NVIDIA NemoClaw ecosystem:
- •NVIDIA OpenShell the underlying sandbox runtime from NVIDIA Agent Toolkit.
- •OpenClaw the always-on AI assistant project this stack is designed to run.
- •NVIDIA Agent Toolkit broader collection of NVIDIA agent infrastructure components.
- •Discord community NVIDIA-hosted community for NemoClaw users and developers.
Getting started:
Minimum hardware: 4 vCPU, 8 GB RAM, 20 GB disk (16 GB RAM and 40 GB disk recommended). Follow the quickstart in the README to install Docker, k3s, and the OpenShell gateway, then push the sandbox image (about 2.4 GB compressed) and start your first OpenClaw agent. Full docs at docs.nvidia.com/nemoclaw/latest.
Limitations:
NemoClaw is alpha software as of May 2026 -- interfaces, APIs, and behavior may change without notice. Memory-constrained machines (under 8 GB RAM) can hit OOM during sandbox image push without configuring swap. The stack is opinionated about running OpenClaw specifically; adapting it for other agent runtimes requires custom blueprint changes. NVIDIA-specific paths mean some features assume NVIDIA infrastructure or NVIDIA-hosted inference endpoints.
Our Verdict
NVIDIA NemoClaw in 2026 is the official reference stack for running OpenClaw always-on AI assistants safely on NVIDIA infrastructure. 20,500 stars in two months, 2,700 forks, Apache 2.0, alpha status, and backed directly by NVIDIA with the broader OpenShell + Agent Toolkit ecosystem behind it.
For Vibe Builders running OpenClaw, NemoClaw is the recommended sandbox stack -- the always-on autonomous nature of OpenClaw assistants makes sandboxing genuinely important, and rolling your own OpenShell setup is error-prone. The guided onboarding handles the Docker, k3s, and gateway plumbing.
For Developers and ML infra teams, NemoClaw is a reference blueprint for the OpenShell sandbox pattern. The architecture docs and component model are useful for designing your own agent sandbox even if you do not run OpenClaw itself.
Skip NemoClaw if you are not running OpenClaw -- the blueprint is opinionated for that specific assistant. Skip if you do not have NVIDIA infrastructure or comfort with k3s + Docker -- the alternative is running OpenClaw less safely or building your own sandbox. For lighter-weight agent sandboxing, plain Docker containers with strict capabilities can work, though without the routed inference and state management features NemoClaw adds.
Frequently Asked Questions
What is NVIDIA NemoClaw?
NVIDIA NemoClaw is NVIDIA's open-source reference stack for running OpenClaw always-on AI assistants more safely inside NVIDIA OpenShell. It installs the OpenShell runtime (part of NVIDIA Agent Toolkit) plus a hardened blueprint, state management, routed inference, and layered protection. As of May 2026 it has 20,500 GitHub stars in two months and is in alpha preview under Apache 2.0.
What is OpenClaw and why does it need a sandbox?
OpenClaw is an always-on AI assistant project. Always-on autonomous agents that can execute commands and access local resources are higher-risk than chat-only assistants. NemoClaw provides a hardened sandbox so the agent runs with limited capabilities, monitored channels, and routed inference. NVIDIA recommends NemoClaw over running OpenClaw bare on a developer machine.
What hardware do I need for NemoClaw?
Minimum: 4 vCPU, 8 GB RAM, 20 GB free disk. Recommended: 4+ vCPU, 16 GB RAM, 40 GB free disk. The sandbox image is about 2.4 GB compressed. On machines with less than 8 GB RAM, configure at least 8 GB of swap to avoid OOM during the initial image push. An NVIDIA GPU is helpful for local inference but not strictly required if you use NVIDIA-hosted endpoints.
How do I install NVIDIA NemoClaw?
Follow the quickstart in the README. The installer handles Docker, k3s (lightweight Kubernetes), and the OpenShell gateway. Then it pushes the sandbox image and starts your first OpenClaw agent. Full step-by-step docs live at docs.nvidia.com/nemoclaw/latest. Be prepared for some setup time on first install while the sandbox image is built and pushed locally.
Is NemoClaw production-ready?
No. NemoClaw is explicitly alpha software as of May 2026, shared by NVIDIA to gather feedback and enable early experimentation. APIs and behavior may change without notice. Do not use it for production workloads yet. NVIDIA welcomes issues and community discussion via GitHub and the Discord channel while the project iterates toward stability.
What is NemoClaw?
NVIDIA NemoClaw is NVIDIA's open-source reference stack for running OpenClaw always-on AI assistants safely inside NVIDIA OpenShell. Layers guided onboarding, a hardened blueprint, state management, and routed inference on top of the OpenShell sandbox runtime.
What license does NemoClaw use?
NemoClaw uses the Apache-2.0 license.
What are alternatives to NemoClaw?
Explore related tools and alternatives on My AI Guide.
Open source & community-verified
Apache-2.0 licensed: free to use in any project, no strings attached. 20,565 developers have starred this, meaning the community has reviewed and trusted it.
Reviewed by My AI Guide for relevance, quality, and active maintenance before listing.
Topics