NVIDIA/NemoClaw

⭐ 20,256  ·  TypeScript  ·  GitHub Repo

Run OpenClaw more securely inside NVIDIA OpenShell with managed inference

1-Sentence Summary

NVIDIA's hardened sandbox for running OpenClaw agents securely with managed inference and cost optimization.

🔥 Key Capabilities & USP

• Hardened Sandbox by Default: Automatically wraps OpenClaw agents in Landlock, seccomp, and network namespace isolation. Solves the pain point of manually configuring secure agent execution environments—no root required, zero configuration needed.
• Guided Onboarding Wizard: Single-command installer (curl | bash) that interactively sets up Node.js, OpenShell, sandbox, inference, and security policies. Eliminates the friction of multi-step setup for developers who want to start experimenting immediately.
• Experimental Model Router: Routes queries to the most cost-efficient model from a configurable pool using a prefill-based accuracy predictor. Directly addresses the pain point of runaway inference costs when running autonomous agents at scale.
• OpenShell-Managed Gateway: All agent communication routes through the OpenShell L7 proxy, keeping API keys and credentials secure from the sandboxed agent. Solves the security concern of exposing sensitive credentials to autonomous code.
• Structured Blueprint Lifecycle: Provides state management and reproducible deployment blueprints. Solves the pain point of managing ephemeral agent state and ensuring consistent deployments across environments.

USP: NemoClaw is the only reference stack that combines NVIDIA OpenShell's managed inference with layered Linux sandboxing (Landlock + seccomp + netns) and a cost-optimizing model router in a single, opinionated setup.

Technical Architecture

Component	Technology	Role
Host	Node.js, npm, Docker	Runs the CLI, installer, and optional Model Router (LiteLLM proxy on port 4000)
Sandbox	Landlock, seccomp, netns	Isolates OpenClaw agent with layered kernel-level security
Gateway	OpenShell L7 proxy	Mediates all traffic between sandbox and external services; protects credentials
Model Router (Experimental)	PrefillRouter encoder (e.g., Qwen3.5-0.8B)	Selects cheapest model meeting accuracy threshold; configured via pool-config.yaml
Data Flow	Sandbox → OpenShell Gateway → Model Router (optional) → NVIDIA API	All agent communication is proxied and secured

Quick Start Guide

Interactive installation (recommended for first-time users):

curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash

Non-interactive installation (for automation):

curl -fsSL https://www.nvidia.com/nemoclaw.sh | NEMOCLAW_NON_INTERACTIVE=1 NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1 bash

Connect to your sandbox and chat via TUI:

nemoclaw my-assistant connect
openclaw tui

Send a single message via CLI:

openclaw agent --agent main --local -m "hello" --session-id test

Onboard with model router enabled:

NEMOCLAW_PROVIDER=routed nemoclaw onboard --non-interactive

Uninstall NemoClaw:

nemoclaw uninstall

Fallback uninstall (if CLI is broken):

curl -fsSL https://raw.githubusercontent.com/NVIDIA/NemoClaw/refs/heads/main/uninstall.sh | bash

Pros, Cons & Use Cases

Pros

• Single-command install with interactive or non-interactive modes—minimal friction to get started
• No root required for sandbox setup; runs entirely in user space
• Hardened by default with three layers of kernel-level isolation
• Cost-optimizing model router that can significantly reduce inference spend
• Multi-platform support: Linux, macOS, Windows WSL2, NVIDIA DGX Spark

Cons

• Alpha software—not production-ready; interfaces may change without notice
• Requires Docker as a prerequisite
• Minimum 8 GB RAM required; OOM risk below that threshold
• macOS and Windows WSL2 have tested limitations and may not support all features
• Experimental model router is still in early development and may produce unexpected results

Who should NOT use this?

• Production teams needing stable, versioned APIs—this is alpha software with breaking changes expected
• Resource-constrained environments with less than 8 GB RAM
• Developers who want full control over their agent sandbox configuration—NemoClaw is opinionated
• Teams not using NVIDIA hardware—while it runs on Linux/macOS/WSL2, the optimizations target NVIDIA ecosystems

Ideal Use Cases

• Security-conscious developers who want to experiment with autonomous agents without exposing their systems
• Researchers evaluating OpenClaw agents in a managed, reproducible environment
• Teams prototyping agent workflows that need cost-aware model routing
• NVIDIA hardware owners (DGX Spark, RTX) wanting a reference stack for managed inference
• DevOps engineers looking for a turnkey sandbox solution for CI/CD agent testing

Community & Activity

With 20,256 stars and an active update as recent as May 2026, NemoClaw has clearly struck a chord with the developer community. This is not a stale project—it's being actively developed by NVIDIA with clear momentum. The star count alone places it among the most popular agent infrastructure projects on GitHub, signaling strong interest in secure, managed autonomous agent deployments. Given the alpha status and rapid iteration, now is an excellent time to evaluate and contribute feedback that will shape the project's direction.