Runs on a private AI box you own
One compact GB10-class device — 128 GB unified memory — runs it on-premise, sold and provisioned by us. No cloud API keys, no data leaving the building.
Run NemoClaw on a private AI box
License: Apache-2.0
NVIDIA's reference stack for running agents sandboxed with routed inference — a natural fit for the NVIDIA GB10-class box it can point at.
What it is
- A reference stack that runs AI agents in a sandboxed "OpenShell" with a routing layer.
- Wraps agents that already speak OpenAI-compatible.
- Published by NVIDIA under Apache-2.0.
NemoClaw is an independent project. BrainOutput has no affiliation, endorsement, or partnership with it — we sell the private hardware you can run it on. Project: NemoClaw.
Cloud API keys vs your private box
The tool is the same either way. What changes is where your prompts and files go, and how you pay.
| Cloud API keys | Your private box | |
|---|---|---|
| Data location | Prompts and files are sent to a third-party API | Everything stays on hardware in your office |
| Cost model | Metered per token — scales with usage, indefinitely | A fixed hardware cost — local inference is unmetered |
| Availability | Vendor uptime, rate limits, and model retirements apply | Runs offline; you control uptime and model versions |
| EU compliance | Data-transfer and processor terms to manage per vendor | EU hardware on your premises — sovereignty is physical, not contractual |
| Model choice | Whatever the vendor exposes | Any open-weight model that fits the box — swap freely |
Indicative hardware start prices are €3 650,22 HT (ASUS) / €4 328,25 HT (Dell); the full sovereign setup is sur devis.
How it connects to your box
NemoClaw's routing layer can direct inference to local models; the agents it wraps use OpenAI-compatible endpoints.
- Route inference to the OpenAI-compatible endpoint on your box.
- Run the wrapped agents against your private model, sandboxed.
How BrainOutput fits
- Our devices are NVIDIA GB10-class hardware — the same family this stack targets.
- We provision the local model endpoint NemoClaw routes to.
Frequently asked questions
- Is NemoClaw from NVIDIA?
- Yes — it is published in NVIDIA's GitHub organization under Apache-2.0. We are not affiliated with NVIDIA; we sell compatible GB10-class hardware.
- Can it use a local model?
- Its routing layer can direct inference to local models, so it can point at the OpenAI-compatible endpoint on your box.
- Why does the NVIDIA box matter?
- The devices we sell are NVIDIA GB10-class systems — the hardware family this stack is designed around.
Run it on a box we set up for you
We size the device, install an OpenAI-compatible model stack, and hand you the endpoint this app points at. The app stays free and yours.