Particle.news
Download on the App Store

Nvidia Ships NemoClaw Blueprints to Run Agents in Governed Production

NemoClaw packages a sandbox runtime with Nemotron models plus the NeMo Agent Toolkit so enterprises can run agents safely on local GPUs.

Overview

  • Nvidia frames an agent as “an LLM and a harness,” meaning the model provides reasoning while the harness supplies loops, tooling, memory and orchestration for production use.
  • NemoClaw has been released as an open collection of blueprints that installs in one step and targets the OpenClaw and Hermes agent frameworks to move prototypes into governed deployments.
  • OpenShell, NemoClaw’s runtime policy layer, sandboxes every agent session, meters resource use, checks permissions at binary/path/method level and logs every allow or deny for forensic audit.
  • Nvidia says it now has full-time engineers contributing to the OpenClaw project and is merging work to address a large PR backlog, signaling deeper upstream involvement in the open ecosystem.
  • NemoClaw routes inference to local Nemotron models by default with optional, policy-controlled routing to frontier models, a design meant to ease enterprise governance and expand demand for on-prem GPUs.