Nvidia said AI Agents are getting easier to use and creators can now run personal AI agents locally on DGX Station desktops.
While DGX Stations are a lot more capable than your typical desktop computer, they are built in a desktop form factor, and now they can run AI agents locally, like a personal computer. Officially, Nvidia calls the DGX Station a “deskside AI supercomputer.”
In a post by Adel El Hallak, vice president of agentic AI at Nvidia, the AI and graphics chip maker announced at Siggraph that super agents have arrived on the desktop.
El Hallak said Nvidia DGX Station is the ultimate deskside supercomputer for the AI era, and with Nvidia Agent Toolkit software, setup takes just three steps, and can be running in roughly 30 minutes.
On DGX Station, Nvidia Agent Toolkit brings together Nvidia NemoClaw, the Nvidia Nemotron 3
Ultra open model, Nvidia Omniverse libraries as agent-accessible tools and skills, and a secure
runtime in a single local system — no internet required.
As workloads scale, developers can connect multiple systems together to serve concurrent users, more agents and bigger models.
This gives creatives and engineers the ability to own their own intelligence, with a system that
comes ready to run locally. The full stack— model, agent, tools — provides a platform for creating and running domain-specific “super agents” that are customized with users’ own data and knowledge.
The open Nvidia Agent Toolkit stack on DGX Station includes:
● Nvidia NemoClaw, open blueprints for building custom autonomous agents, packaging
the model, harness and runtime together as a starting point for teams building specialized,
domain-specific agents.
● Nvidia Nemotron 3 Ultra, a frontier 550-billion-parameter open model, is optimized to run
on DGX Station GB300 systems and serves as the model layer that teams can customize
for their own domains.
● Nvidia Omniverse libraries extend agent skills into physics simulation and 3D asset
workflows, giving creative and engineering professionals tools that go well beyond
general-purpose agent capabilities.
● NvidiaOpenShell, the open source secure runtime, keeps agents sandboxed and
governed according to defined policies for how agents interact with tools, systems and
data.
● Nvidia GB300 Grace Blackwell Ultra Desktop Superchip delivers data-center-level
performance from the desk on DGX Station, with up to 20 petaflops of FP4 AI compute
and 748GB of coherent memory to run large models such as Nemotron Ultra.
● Nvidia ConnectX-8 SuperNIC delivers up to 800GB/s of bandwidth in DGX Station,
delivering extremely fast, efficient network connectivity, and supports linking up to two
DGX Stations to further scale model capacity and performance.
Harness Efficiency at Scale
For teams running agents at scale the economics shift fundamentally on DGX Station. Nemotron 3 Ultra, tuned for an open harness, delivers leading-edge performance without the per-token cost after the hardware purchase, so users build once and can run as much as they need.
Nvidia has announced a blueprint for integrating Nvidia Omniverse libraries in Blender — giving NemoClaw agents callable RTX sensor simulation and physics tools to prepare 3D scenes for physical AI workflows.
On DGX Station, designers and engineers can run the core pieces of that workflow — frontier
model, open harness, secure runtime, 3D tools — in one box, all connected and deployable
through an open blueprint. And Nvidia AI Agent tookit now has Omniverse libraries.
Frontier models can orchestrate NemoClaw as a specialized sub-agent, delegating domain-
specific work to an agent running locally on DGX Station, with direct access to Omniverse tools and Blender.
LangChain tuned its Deep Agents harness for Nemotron 3 Ultra, giving designers and engineers a production-ready path to benchmark-leading agentic performance at a fraction of the cost.
Nous Research fine-tuned Nemotron 3 Ultra for its Hermes Agent harness and adopted it for production workloads — a direct demonstration of the value of owning intelligence. Tuning the
model for a developer’s stack enables agents that are both faster and more capable for specific
domains. Hermes Agent has also added Blender to its Model Context Protocol catalog, letting
teams activate Blender directly from their agent — a live example of a tool-using NemoClaw agent that can run on DGX Station.
For teams running OpenClaw, this stack extends what’s possible — bringing Nemotron 3 Ultra,
Omniverse tools and local inference on DGX Station into an environment where OpenClaw’s
persistent, long-running agents can act on them continuously.
Develop and Deploy Quickly With New Playbooks
Two new playbooks are available now to help developers build and run agents out of the box with NemoClaw and dual-node deployments:
● “Connect Two DGX Stations for Distributed Workloads” guides users on how to run powerful models with scalable performance connecting up to two DGX Station systems.
● “Run NemoClaw With a Local LLM / On Dual DGX Station” provides a walkthrough on how to set up NemoClaw on DGX Station, enabling out-of-the-box agents with frontier intelligence with Nemotron Ultra.
Nvidia DGX Station is built and available to order from ASUS, Dell Technologies, Exxact,
GIGABYTE, HP, MSI and Supermicro.