Nvidia unveils Agent Toolkit and NemoClaw for fast, secure deployment of AI agents

At GTC 2026, Nvidia bet not only on new models but also on the infrastructure to put them to practical use. The company introduced Agent Toolkit and the NemoClaw stack, two open-source tools meant to make AI agents easier to launch, safer, and closer to real business scenarios.

Two new tools

Nvidia positions Agent Toolkit as a base layer for enterprise agent systems. The idea is that developers should not rebuild every project from scattered libraries, but get a common set of components for connecting agents to data, tools and workflows. The platform is designed to work alongside existing frameworks, not replace them, so it can be added to a current stack without a full migration.

According to Nvidia's documentation, the Toolkit works as a unifying layer on top of popular agent frameworks, including LangChain, LlamaIndex, CrewAI, Semantic Kernel and Google ADK. It has built-in profiling, tracing, response quality evaluation and support for the MCP and A2A protocols. In practice, companies are being offered not just a library for running an agent, but a set of tools for observability, debugging and gradual scaling of such systems inside the enterprise.

What's inside the stack

If Agent Toolkit handles building and operating agent applications, NemoClaw addresses a different problem: how to quickly bring up a ready-made autonomous assistant on your own hardware without losing control over what it does. The stack is built around OpenClaw, but adds Nvidia Nemotron models and a new runtime, OpenShell, which handles security policies, privacy and access control. In effect, NemoClaw turns OpenClaw into a more manageable package for local or dedicated deployment.

Nvidia stresses separately that the stack can be deployed with a single command and can run at different levels of infrastructure, from PCs with GeForce RTX GPUs and workstations to DGX Spark, DGX Station and servers with Nvidia accelerators. The company also cites optimization for more powerful Vera Rubin systems and support for clustering up to four DGX Spark units in one configuration.

• installation of OpenClaw, Nemotron and OpenShell in a single launch flow
• network and file system isolation for agent actions
• policies for outbound requests and data access
• local inference, so data does not have to be sent to the cloud
• scaling from RTX PCs to servers and compact DGX systems

Notably, Nvidia does not sell NemoClaw as a magic "launch the agent and forget it" button. It is rather a reference stack for those who want an always-on assistant with clear boundaries: where it can read files, which network calls it is allowed to make, and which actions require separate approval. The documentation labels the project alpha, so for now it is an early tool for developers rather than a mature, out-of-the-box enterprise product.

Partners and ecosystem

The announcement came with a long list of partners that are ready to use Agent Toolkit in their products. Nvidia named Adobe, Salesforce, SAP, ServiceNow, Siemens, CrowdStrike, Atlassian, Palantir and several other companies. The logic of the partnerships is clear: Nvidia wants its agent stack to become not an experiment for enthusiasts but an infrastructure foundation for enterprise software, where agents live long, work with internal data and are embedded in real business processes.

The company singled out its collaboration with Adobe. There, Agent Toolkit is seen as the basis for hybrid, long-running agent scenarios in creative, productivity and marketing applications. In parallel, Adobe plans to integrate CUDA-X and Omniverse more deeply, and training of next-generation models will rely on Nvidia's computing platform. For Nvidia itself, this is a way to establish itself not only as a GPU supplier but also as the owner of the software layer over the agent economy.

"OpenClaw has opened the next frontier of AI to everyone," said

Jensen Huang.

Another strategic move is the launch of the Nemotron Coalition. Its members are Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam and Thinking Machines Lab. These companies will jointly develop open frontier models on shared DGX Cloud infrastructure, to avoid duplicating the basic training stages and move faster to industry specialization. Nvidia says directly that the results of this work will form the basis of the future Nemotron 4 family.

What this means

The AI market is shifting from a race between individual models to a race between agent platforms. Nvidia wants to hold both positions at once: supply the models, the hardware and the management layer on top of them. If the Agent Toolkit and NemoClaw approach takes hold, the company could become for enterprise AI agents what CUDA once became for GPU computing: the standard around which the whole ecosystem is built.