AgentBOX is a pre-configured AI assistant on hardware you own. The agent, the memory, the companion app, and the first skill all ship ready — plug it in, open the app, and start talking. Pick the configuration that fits your job; the same membership works across all of them.
An assistant that learns you, on hardware you own.
A single SKU at launch — the AgentBOX appliance. It ships with the assistant, the companion app, the MailBOX skill, and the OpenClaw conversational front-end already configured. Plug it in and start talking — through the app, your email, or any of the messaging bridges.
On the 8GB Jetson, AgentBOX keeps one skill active at a time and hot-swaps between the ones you've enabled (MailBOX, OpenClaw, more as they ship). Need multiple skills truly concurrent? That's multi-skill hosting — it lives on qualified hardware (T3+), below.
Your membership says how many skills you're entitled to. Your hardware says how many host at once. And your box's architecture — not its price — decides how hard the OpenClaw sandbox is. Tier tells you capability; chip tells you hardening.
Launch is the AgentBOX appliance on the T2 Jetson. Below it, run the same agent on hardware you already own or a low-cost Pi; above it, the ladder is NVIDIA-native the whole way up — same software, more headroom — with a single x86 option when you need the hardened OpenClaw sandbox. Tap any tier for the detail.
The sovereignty path — and the cheapest. Install the full thUMBox stack — local model server, vector store, workflow engine, and dashboard — on a machine you already own: a desktop, mini PC, laptop, or Apple-silicon / NVIDIA box. One command brings the containers up; the model downloads once and stays local. It runs on the free Community plan, no card required. Capability scales with the machine — 8 GB runs a single pack comfortably, 16 GB and up hosts more than one at once, and on an x86 box OpenClaw can use the hardened OpenShell sandbox. Your metal, your models, your tokens — nothing phones home.
The cheapest path to a dedicated, always-on node. A Raspberry Pi 5 — or any equivalent single-board computer — runs thUMBox on ARM64 at a few watts, around the clock, for the price of a nice dinner. Inference is CPU-paced and sized to small models, so this is the floor of the ladder: think a private on-device chatbot and light email triage rather than heavy drafting. What you trade in speed you get back in cost and footprint — a tiny box on a shelf that's wholly yours, with no per-token meter and nothing leaving your network.
Available now. The Jetson Orin Nano Super fits 67 TOPS into a 15-watt envelope — purpose-built silicon that runs a 4B-class model on-device at ~18 tokens/sec. AgentBOX ships as a single SKU with both the MailBOX skill (inbox triage and drafting in your voice) and the OpenClaw conversational front-end (companion app, Telegram, WhatsApp, web) configured out of the box. The 8 GB envelope keeps one skill active at a time; a Plus membership hot-swaps between your entitled skills with the agent kept warm. OpenClaw runs with Docker-level isolation and ships disabled by default on this hardware. Every AgentBOX includes a year of Base membership.
The natural step up, on the same NVIDIA stack as your launch box. The Jetson AGX Orin brings 64 GB and up to 275 TOPS — the headroom to host several packs concurrently and run larger models, with no change to the software you already know. OpenClaw stays Docker-level here: ARM64 keeps the same sandbox posture as T2, with far more capability. This is the primary T3, and the box receptionBOX re-platforms onto.
The x86 path — chosen when hardened isolation is a requirement, not a nice-to-have. The Mac mini M4 (24 GB unified) delivers the same T3-class multi-pack capability and adds NemoClaw's kernel-level OpenShell sandbox: Landlock, seccomp, and network namespaces wrapped around the agent runtime. Hardening tracks the chip, not the price — this is the deliberate choice when OpenClaw needs to be locked down tighter than Docker-level.
On the roadmap. A multi-GPU workstation with 64 GB and up runs 14–30B models and orchestrates several agents at once — enough for a small team to share a single box. The x86 path carries NemoClaw's OpenShell sandbox by default. Built for the moment one operator's box becomes a team's shared brain.
On the roadmap. Rack- or fleet-scale hardware with 128 GB and up runs 30–70B models under fleet-coordinated policy — intelligence that spans an organization while still living on metal you own. NemoClaw enforces sandboxing fleet-wide. The top of the ladder: data-center-class capability with zero data-center dependency.