NVIDIA local agent tutorial

NVIDIA published a step-by-step guide on April 17 for building a local AI agent that stays on a user’s own machine instead of sending work to the cloud. (developer.nvidia.com) The tutorial pairs OpenClaw, a “local-first” agent, with NVIDIA NemoClaw, an open-source reference stack that installs OpenShell guardrails and open models with a single command. NVIDIA says the walkthrough runs on its DGX Spark system and connects the agent to Telegram for remote access. (developer.nvidia.com) In plain terms, an AI agent is a bot that can keep running, read files, use tools, and take actions across multiple steps instead of answering one prompt at a time. NVIDIA’s materials say that kind of autonomy raises risks around data exposure and code execution when the agent is not isolated. (developer.nvidia.com) NemoClaw is NVIDIA’s packaging layer for that problem. Its product page says it adds policy-based privacy and security controls through OpenShell, while OpenClaw supplies the assistant that can run continuously and use local context from files and apps. (build.nvidia.com, build.nvidia.com) The DGX Spark playbook says the installer sets up Node.js, OpenShell, and the NemoClaw command-line tool, then walks the user through creating a sandboxed agent with Ollama and NVIDIA’s Nemotron 3 Super model. The same guide says the finished setup can be reached through a web dashboard or a Telegram bot. (build.nvidia.com) NVIDIA’s instructions are explicit about the hardware and software plumbing. On DGX Spark, the guide requires Docker, the NVIDIA container runtime, Ollama listening on all interfaces, and a local download of Nemotron 3 Super 120B, which the instructions list at about 87 gigabytes. (build.nvidia.com) The company is also explicit that this is not a production recipe. The NemoClaw developer guide labels the software “alpha,” says it has been in early preview since March 16, 2026, and warns that interfaces and runtime behavior can break between releases. (docs.nvidia.com) The security warnings are as prominent as the setup steps. NVIDIA says users should run the demo on a fresh device or virtual machine with no personal data, and its OpenClaw guide warns that the agent can access files, execute commands, and connect to external services. (build.nvidia.com, build.nvidia.com) That framing puts the tutorial closer to a controlled lab exercise than a plug-and-play consumer product. NVIDIA’s own documentation says the aim is a sandboxed, always-on assistant with local inference and tighter policy controls, not a guarantee that autonomous agents are safe by default. (developer.nvidia.com, build.nvidia.com)