NVIDIA launches open-source “NemoClaw” to run always-on AI agents more safely

NVIDIA has introduced NemoClaw, a new open-source stack designed to help developers run autonomous, always-on AI agents with stronger security and privacy controls. The project is currently available in early alpha.

Adding a safety layer to the AI agent stack

NemoClaw builds on top of OpenClaw, positioning itself as a security-focused extension rather than a standalone framework. Its goal is to make autonomous agents—capable of writing code, calling APIs, and operating continuously—more controllable in real-world environments.

At the core is NVIDIA OpenShell, a runtime that enforces policy-based guardrails across the entire execution lifecycle. This includes:

• Blocking unauthorized outbound network requests
• Restricting filesystem access outside sandboxed directories
• Preventing privilege escalation and unsafe system calls
• Routing all model inference through controlled gateways

Instead of trusting the agent, the system enforces constraints at the infrastructure level.

One-command setup, sandboxed by default

NemoClaw is designed for fast onboarding. Developers can install and initialize the stack with a single command, which sets up the runtime, sandbox, and agent environment.

Once deployed, agents run inside isolated containers governed by strict policies. Any attempt to access unapproved external resources can be intercepted and surfaced for user approval, introducing a human-in-the-loop control mechanism.

Hybrid architecture: local execution, managed inference

The system uses a hybrid model. Agents execute locally—handling logic, tool use, and automation—while inference is routed through NVIDIA-managed infrastructure.

Models such as NVIDIA Nemotron are supported, with requests passing through OpenShell rather than being sent directly from the agent. This design enables auditing, filtering, and policy enforcement on every inference call.

Built for persistent, always-on agents

NVIDIA is clearly targeting a future where AI agents run continuously rather than on demand. NemoClaw is optimized for dedicated compute environments, including:

• RTX-powered PCs and laptops
• Professional workstations
• DGX-class systems

These setups allow agents to operate 24/7, performing tasks such as coding, monitoring, and automation without interruption.

Still early: powerful, but not production-ready

NVIDIA describes NemoClaw as alpha software, emphasizing that the project is still evolving:

• APIs and interfaces may change
• Some features are incomplete
• Setup may require manual workarounds

For now, the focus is on experimentation and developer feedback rather than production deployment.

Why this matters

AI agents are rapidly shifting from chat interfaces to autonomous systems that can act, decide, and execute. That shift introduces a new category of risk: agents with real capabilities but insufficient control.

NemoClaw reflects a broader industry trend toward “governed autonomy”—where agents are powerful, but their actions are tightly constrained by policy and infrastructure.

Instead of asking what agents can do, tools like NemoClaw focus on defining what they should be allowed to do—and enforcing it by default.