Run OpenClaw on a private AI box
A sovereign personal AI agent whose data, files, and automations never leave the company.
OpenClaw is an open-source personal AI agent that can read your files, run shell commands, automate the web, and talk to your team through the chat apps they already use. Out of the box, most people wire it to a cloud provider's API key — which means every prompt, every document, and every action it takes flows through someone else's servers.
It does not have to. OpenClaw is model-agnostic: it will just as happily talk to a private model running on hardware you own. This page is about that setup — OpenClaw pointed at a private AI box that sits in your office, so the agent is genuinely yours end to end.
What OpenClaw is
OpenClaw is an independent open-source project — a local-first personal AI agent released under the permissive MIT license, with a large and fast-growing community (on the order of a quarter of a million GitHub stars by early 2026). Because it is open source and model-agnostic, you are never locked into one vendor's model or one company's cloud.
Its reach is what makes it useful as a day-to-day agent, and also what makes where it runs matter so much:
- Model-agnostic, bring-your-own-model: point it at a cloud API key, or at a fully local model running on your own infrastructure — the same agent, a different backend.
- Meets your team where they already are: it connects to chat apps such as WhatsApp, Telegram, and Signal, so the agent is reachable without a new tool to learn.
- Does real work, not just chat: a large catalogue of AgentSkills (100+) gives it shell access, file operations, and web automation — the same capabilities that make sovereignty over its backend matter.
- Permissive MIT license: you can self-host, modify, and run it commercially at no licence cost.
To be precise, in the same way we are precise about OpenCode and LibreChat elsewhere on this site: OpenClaw is an independent open-source project, and we are not affiliated with it, its stewarding foundation, or its sponsors. This is not a partnership or an endorsement in either direction. What BrainOutput provides is the private hardware and the tailored local models that OpenClaw can run against — not OpenClaw itself, which you are free to run however you like. OpenClaw.
Why a private box beats cloud API keys for OpenClaw
An agent with shell access, your files, and your chat history is exactly the workload you least want flowing through a third party's servers. Running OpenClaw against a private box changes the trade on every axis that matters:
| OpenClaw on cloud API keys | OpenClaw on a private AI box | |
|---|---|---|
| Where your data goes | Every prompt, file, and action flows to a third party's servers. | Prompts, files, and automations stay on hardware in your building — nothing leaves the company. |
| Cost model | Metered per token; an always-on agent that reads and acts continuously runs up an open-ended bill. | A fixed hardware cost you own — no per-token meter on the agent's day-to-day work. |
| Availability | Depends on a provider's uptime, rate limits, and model-deprecation schedule. | Always-on local models under your control; no rate limit, no surprise deprecation. |
| EU compliance | Cross-border data flows and processor agreements to manage under the GDPR and the EU AI Act. | On-premise by construction — the strongest posture for GDPR and EU AI Act obligations. |
| Integration | A separate account and key, disconnected from the models your team already runs. | Runs against the same private LLM your team already uses — one sovereign endpoint for everything. |
None of this requires giving anything up on capability: a modern GB10-class device runs large open-weight models locally, so the agent behind OpenClaw can be a genuinely capable model, not a toy.
A GB10-class box is a natural, vendor-acknowledged host
This is not only our read of the fit. ASUS advertises its Ascent GX10 — a GB10-class desktop AI supercomputer with 128 GB of unified memory — as "Supports OpenClaw" in its official product specifications, and the same wording appears across major retail listings (Amazon, Newegg). The hardware vendor markets this exact class of device as an OpenClaw host.
That is precisely the class of device BrainOutput sells and provisions. So the sovereign-box route for OpenClaw is not a workaround — it is the use the device makers themselves advertise, with the added value that we set up the private models for you.
How BrainOutput fits
We do not make OpenClaw, and we do not resell it — it is free and open source, and you run it yourself. What we provide is the private AI box it points at, ready to be a good backend for it.
Concretely, that means we sell and provision the sovereign device, then tailor the local models that run on it to your work — so that when you connect OpenClaw, there is a capable, private model waiting on the other end, entirely inside your company.
- We sell and provision the GB10-class device — the same hardware class ASUS markets as OpenClaw-capable.
- We install and tailor the local models it serves, sized to your workload, so OpenClaw has a genuinely capable private backend.
- It runs against the same private LLM the rest of your team already uses — one endpoint, one sovereign machine.
- Full setup — device, models, tailoring, installation, support — is scoped and priced per assessment (sur devis); OpenClaw itself stays free and yours.
Frequently asked questions
- Is BrainOutput affiliated with OpenClaw?
- No. OpenClaw is an independent open-source project (MIT-licensed), and we are not affiliated with it, its foundation, or its sponsors. We sell and provision the private hardware and the local models it can run against — we do not make or resell OpenClaw, which you run yourself.
- Can OpenClaw really run without any cloud API key?
- Yes. OpenClaw is model-agnostic and bring-your-own-model: instead of a cloud provider's key, you point it at a private model served on your own device. The agent behaves the same way — only the backend changes, and with it, where your data lives.
- Why does it matter where OpenClaw's model runs, if the agent is the same?
- Because OpenClaw is not just a chatbot — with its AgentSkills it reads your files, runs shell commands, and automates the web. On a cloud key, all of that context passes through a third party. On a private box, it stays inside your company, which is the difference between a contractual privacy promise and a physical one.
- Which device do you provide for this?
- A GB10-class device with 128 GB of unified memory — the same class ASUS advertises as "Supports OpenClaw". We sell two variants and provision them with a working private model stack; the memory-fit maths for the models they run is on our on-premises model pages.
- What does it cost?
- OpenClaw itself is free and open source. The private box, the local models, tailoring, installation, and support are scoped and priced per assessment (sur devis) — we size the setup to your workload rather than quoting a number that ignores it. Request an assessment and we will scope it with you.
Make OpenClaw yours, end to end
Request an assessment and we will scope the private box and the local models OpenClaw would run against — for your workload, your compliance needs, and your team.