Most AI assistants live on someone else's server. Your emails, your browsing, your schedule — all processed in a cloud you don't control, behind a subscription that never ends.
ClawBox flips that. It's a small, plug-and-play box built on an NVIDIA Jetson Orin Nano Super (67 TOPS) that runs a real AI agent locally — on hardware you own, on your own network.
What it does, quietly in the background:
- Browser tasks — research, form-filling, monitoring pages, pulling data.
- Messages — triage and draft email, keep your inbox under control.
- Schedules — set tasks before bed, wake up to them done.
The part people miss: you're not locked to the box's chip. Run private models locally, or point ClawBox at your own OpenAI or Anthropic key when a job needs a frontier model — your choice, per task. So it keeps up as models improve, instead of going out of date.
The hardware is €549. Local workloads stay on your machine; optional AI plans and cloud providers are available when you choose them.
What “Local-First” Actually Means
Local-first is not a promise that every workflow is offline. The OpenClaw runtime, memory, local files, workflow rules, and a suitable local model can run on the ClawBox itself. Email, messaging platforms, web browsing, cloud models, and remote services still need a network connection and send the data required for those operations.
That distinction gives you a useful control point. A private document can be summarized with a local model. A difficult reasoning task can be routed to an optional cloud model after you decide the context is appropriate to share. A browser workflow can run from your own logged-in session instead of a disposable remote browser, while the websites you visit still receive normal web traffic.
Start With Work You Can Verify
The safest first automations are reversible and easy to inspect:
- Summarize an inbox or folder without changing anything.
- Draft replies for human review instead of sending automatically.
- Monitor a page and notify you when something changes.
- Produce a daily brief with links back to every source.
Add write access only after the read-only version is reliable. For consequential actions, keep approval gates, logs, rate limits, and a kill switch. The goal is not maximum autonomy on day one. It is an assistant you can trust with progressively more useful work.
ClawBox is the guide in that process, not the hero. You choose the work, permissions, models, and privacy boundary; the dedicated hardware keeps the system available to carry out those choices.
Your AI, on your terms.