Nvidia Corp. unveiled NemoClaw and Agent Toolkit at its annual GTC developer conference, introducing open-source solutions to strengthen AI agent performance and security.
NemoClaw enhances OpenClaw, a consumer-run AI agent that automates tasks like file organization and online research. It integrates two core components: OpenShell and Nemotron. OpenShell runs agents in secure sandboxes, limiting file and network access to protect privacy. Developers can customize policies using YAML rules, some of which are hot-swappable without restarting agents.
Nemotron, a suite of AI models optimized for text generation and graph analysis, powers the intelligence layer in NemoClaw. Agents can route queries through Nemotron or integrate with cloud-based large language models while using the Privacy Router to block sensitive data from being transmitted.
The new Agent Toolkit enables developers to build custom AI agents using a blueprint called AI-Q, based on a hybrid architecture. It combines cost-efficient Nemotron models for research with high-end frontier models for orchestration-cutting processing costs by over 50%.
Nvidia reports its AI-Q-powered agent topped the DeepResearch Bench and DeepResearch Bench II leaderboards. The toolkit will integrate with platforms from Adobe, Red Hat, Box, Cadence, and LangChain.
“Employees will be supercharged by teams of frontier, specialized and custom-built agents they deploy and manage,” said CEO Jensen Huang. “The enterprise software industry will evolve into specialized agentic platforms.”