NVIDIA releases NemoClaw and Agent Toolkit; OpenShell sandbox runtime now open-source for enterprise agentic AI

NVIDIA announced NemoClaw, an open-source security stack for autonomous AI agents, and an expanded Agent Toolkit alongside the open-source OpenShell runtime. NemoClaw installs in a single command and wraps OpenClaw agents with enterprise-grade privacy and security controls: kernel-level sandboxing (seccomp, Landlock), filesystem restrictions, network policies, and a Privacy Router that intelligently routes queries between local Nemotron models and cloud inference endpoints based on data sensitivity. The stack targets the gap between agent prototypes and production-grade, policy-governed deployments.

The toolkit comprises NVIDIA Nemotron (open models for local inference), NemoClaw (blueprints for safe agent behavior), and OpenShell (policy-governed runtime). Box, Adobe, Salesforce, Dell, Cisco, and LangChain are early integration partners. On the capability front, NVIDIA BioNeMo Agent Toolkit demonstrates domain-specific agents for life sciences: molecular screening workflows that previously took days compress to minutes, and genomics analysis tasks are now executable within agent-driven discovery loops. The Nemotron Coalition (Mistral AI, Perplexity, Cursor, LangChain) is co-developing Nemotron 4, a base model optimized for agentic use cases.

For production agentic systems, NemoClaw addresses a concrete problem: enterprises need autonomous agents that can reason, take action, and integrate with existing tools, but cannot expose sensitive data, customer records, or proprietary code to external APIs. The Privacy Router solves this by keeping sensitive queries local while routing non-sensitive work to frontier models. Network policies are default-deny; outbound connections require pre-approval or real-time authorization. Audit logging captures every action.

Architects deploying agents should view this as the security and operational foundation for production agentic AI. The single-command install and policy-based approach (YAML-driven) signals NVIDIA's confidence in a multi-agent, long-running, always-on future. The fact that Jensen Huang positioned NemoClaw as an OS-for-agents answer ("OpenClaw is the operating system for personal AI") indicates this layer will be foundational to agentic computing, similar to how Docker became foundational to containerization.