Nvidia launches NemoClaw, Agent Toolkit to enhance AI agents

Nvidia Corp. today introduced a set of open-source tools designed to enhance the capabilities of artificial intelligence agents.

The projects made their debut at the company’s annual GTC developer conference.

Nvidia’s first new tool, NemoClaw, improves the safety and output quality of OpenClaw. OpenClaw is a popular open-source AI agent that consumers can run on their personal computers. It’s capable of automating multi-step tasks such as organizing locally-stored files and performing online research.

NemoClaw comprises multiple components. The first is OpenShell, a tool that adds privacy and cybersecurity guardrails to OpenClaw. It does so by running OpenClaw in a sandbox that prevents it from accessing files not necessary to the tasks it performs. The tool also limits the agent’s network access.

Developers can customize OpenShell by writing configuration rules in the YAML syntax. For example, a software team could create a policy that permits network connections between an agent sandbox and a cloud-hosted AI tool. Some YAML rules are hot-swappable, which means that they can be changed without restarting the affected agent.

The other component of NemoClaw, the new open-source tool that OpenShell powers, is an existing Nvidia project called Nemotron. It includes more than a half dozen AI models optimized for tasks such as generating text and analyzing graphs.

Users can configure their NemoClaw-powered agents to route prompts to Nemotron algorithms. According to Nvidia, it’s also possible to integrate agents with cloud-hosted large language models. A component known as a Privacy Router ensures that sensitive data isn’t sent to those LLMs.

Nvidia says that NemoClaw makes it possible to install both OpenShell and Nemotron with a single command. Besides OpenClaw, NemoClaw can also power other AI agents that use models from providers such as OpenAI Group PBC and Anthropic PBC.

“Employees will be supercharged by teams of frontier, specialized and custom-built agents they deploy and manage,” said Nvidia Chief Executive Officer Jensen Huang. “The enterprise software industry will evolve into specialized agentic platforms, and the IT industry is on the brink of its next great expansion.”

OpenShell and Nemotron, the two core components of NemoClaw, also ship with another open-source project that Nvidia debuted today. Agent Toolkit enables developers to build custom AI agents that automate multi-step tasks. Like NemoClaw, it runs agents in isolated OpenShell sandboxes to boost security.

Agent Toolkit includes a software blueprint called the AI-Q that is designed to speed up development projects. According to Nvidia, software teams can use it to build search agents that automate tasks such as sifting through business documents for specific data points. The chipmaker used AI-Q to develop an agent that topped two AI leaderboards called DeepResearch Bench and DeepResearch Bench II.

The blueprint is based on what Nvidia calls a hybrid architecture. AI-Q uses Nemotron algorithms to perform research and a pricier frontier model to orchestrate those algorithms. According to Nvidia, that approach can cut query processing costs by more than half.

The chipmaker is currently working with a long list of partners to integrate Agent Toolkit with their products. Adobe Inc., IBM Corp.’s Red Hat unit, Box Inc. and Cadence Design Systems Inc. are among the early adopters. Startup LangChain Inc, in turn, will integrate Agent Toolkit with its widely-used AI application development platform of the same name.

Image: Nvidia