NVIDIA pitches NemoClaw blueprints for open agent stacks
NVIDIA is pitching its NemoClaw blueprints as a way to run existing open-source agents on its own stack, which bundles Nemotron models, optimization tools and a governed runtime, without rewriting them.
NVIDIA laid out on its technical blog how its NemoClaw open blueprints fit into an agent stack. The post is an explainer, not a launch announcement.
The company says NemoClaw is neither a harness nor a runtime. It describes it as a set of reference blueprints that build a stack around a harness the developer has already chosen. Three harnesses are supported today: OpenClaw, Nous Research's Hermes Agent and LangChain Deep Agents.
Each blueprint combines four components:
- a version of the harness tuned for NVIDIA's Nemotron models
- Nemotron open models
- NeMo optimization tools, including NeMo Relay for tracing and improving agent runs, plus routing
- NVIDIA OpenShell, a runtime that controls agent access to files, networks, credentials and tools
The only performance claim in the post involves LangChain. According to NVIDIA, LangChain tuned Deep Agents for Nemotron 3 Ultra by changing prompts, tools and middleware, without retraining the model. NVIDIA says this brought the model to "near-frontier performance" at about 10x lower cost per run than leading closed models. The post does not name the benchmarks, the comparison models or the cost method. The tuned model profile is available through LangChain on GitHub, and the Deep Agents blueprint packages those changes.
NVIDIA also points to a "Chief of Staff" recipe in the NemoClaw community repository. It is a memory-driven Hermes profile in which OpenShell keeps credentials for inference and MCP connections outside the sandbox, and NeMo Relay exports session traces for audit.
The post does not cover pricing, availability terms or independent evaluations.