NemoClaw
Run agents like Hermes and OpenClaw more securely inside NVIDIA OpenShell with managed inference
A secure sandbox for running autonomous AI agents locally, with Linux kernel-level isolation to prevent agents from accessing your files or network without permission.
NVIDIA NemoClaw is an open-source reference stack — essentially a pre-configured setup guide and toolset — for running OpenClaw AI assistants in a more secure and controlled way. OpenClaw is a type of always-on autonomous AI agent, meaning it can run continuously in the background and take actions on your behalf. NemoClaw wraps OpenClaw inside NVIDIA's OpenShell runtime, which adds a protective sandbox around the agent so it cannot freely access your system or network in unintended ways.
The core problem NemoClaw solves is that running autonomous AI agents on your own hardware is risky if the agent has unconstrained access to your files, network, and processes. NemoClaw addresses this by applying Linux kernel-level security techniques (Landlock filesystem restrictions, seccomp system-call filtering, and network namespace isolation) to confine the agent. It also handles guided setup, state management, secure messaging between the agent and the outside world, and routing of AI inference requests — either to NVIDIA's hosted model endpoints or to a local model served via Ollama.
Installation is a single curl command that runs a setup wizard, after which you interact with the agent through a terminal interface or command-line messages. An experimental model router feature automatically picks the cheapest AI model capable of handling each query, rather than sending everything to a large expensive model.
NemoClaw is intended for developers and researchers who want to experiment with autonomous AI agents locally without giving those agents free rein over their machine. It requires Node.js 22+, Docker, and runs on Linux, macOS (Apple Silicon), and Windows via WSL2. The project was in early alpha as of early 2026.
Where it fits
- Run autonomous AI agents on your local machine without them accessing files or network outside a sandbox.
- Test and experiment with always-on AI assistants that can take actions on your behalf safely.
- Route AI queries to the cheapest capable model automatically instead of always using expensive large models.
- Set up a secure autonomous agent workflow with guided installation and terminal-based interaction.