thUMBox
Bring Your Own Hardware

The same agent. Your metal.

Don't want an AgentBOX appliance? Run the same self-learning assistant on hardware you already own — a desktop, a mini PC, even the old phones in your drawer. You handle the setup; you get full control. This is the sovereignty path, free on Community.

Install on your machine Read the docs ↳ Runs on the free Community plan. No card required.
Why this matters
Stop renting your own intelligence.

The cloud AI model is simple: they own the compute, so they own you. Every token is metered, every prompt is logged, the price moves whenever they like, and the door can close without notice. Local inference flips the whole arrangement.

The data-center deal
  • ✕Metered by the token. Costs scale with use and rise on their schedule, not yours.
  • ✕Your data is their training set. Every prompt may be logged, retained, and learned from.
  • ✕Models change under you. The model you built on can be deprecated or quietly swapped.
  • ✕Access is a privilege. Rate limits, policy shifts, or an account flag can cut you off.
  • ✕Latency lives elsewhere. Every request makes a round trip to someone else's building.
Bring your own hardware
  • ✓Fixed cost. Zero per-token. You paid for the machine once. Inference is free after that.
  • ✓Your data never leaves. It runs on your metal, on your network. Nothing to log because nothing leaves.
  • ✓You pin the model. Choose your weights, freeze them, swap them on your terms. They don't move unless you move them.
  • ✓No one can revoke you. There's no account to flag. The box answers to you.
  • ✓Inference is on your desk. No round trip. The intelligence is where you are.
How it works
Three steps to a box you own.

thUMBox is a local-first agent runtime — Hermes for tool-calling, gbrain for persistent memory, a workflow engine, and a dashboard, all running in containers on your machine. Installing it on your own hardware is the same software we ship pre-loaded, minus the box.

01
Bring a machine
A recent workstation, mini PC, or NVIDIA / Apple-silicon box you already own. If it can run a 4B model locally, it can run thUMBox.
02
Run the installer
One command pulls the container stack — ollama, vector store, workflow engine, dashboard — and brings it up on your hardware. Your model downloads once and stays local.
03
Own it
Open the dashboard, point it at your accounts, and go. Build your own packs with the open SDK. Nothing phones home. No subscription required to keep it running.

What your machine needs

Memory
8 GB minimum
16 GB+ to host more than one pack at once
Acceleration
NVIDIA GPU or Apple silicon
CPU-only works for light triage; GPU strongly recommended
Storage + OS
~30 GB free · Linux or macOS
Docker-capable; NVMe recommended for the model + index

The full platform sits around 5.5 GB of working memory with a 4B model loaded — so an 8 GB machine runs a single pack comfortably, and anything you'd call a modern desktop or gaming rig has room to spare. Bigger machine, bigger models, more packs hosted at once.

The drawer full of compute
That old phone isn't dead. It's a node.

Everyone has them — the drawer of retired iPhones and Androids, perfectly good silicon gathering dust. Each one is a low-power computer with a real NPU. thUMBox turns yesterday's phone into a private inference node you already paid for. No new hardware. No data center. Pure reclaimed compute.

01
Wake it up
Pull a retired iPhone or Android out of the drawer, factory-reset it, and put it on your network. A phone from the last several years has the NPU to run a small local model.
Install thUMBox
02
Grab the app — App Store, Google Play, or side-load. It brings up the local runtime and dashboard on the device. The model downloads once and lives on the phone.
03
Put it to work
Mount it on a shelf, plug it in, and it runs. On-device triage and drafting, 24/7, on hardware that cost you nothing more. Old phone, new job.
iPhone · iPadA-series / M-series · Neural Engine
AndroidSnapdragon / Tensor · NPU
However much the phone can do

Older or smaller phones won't run the full platform — and that's fine. What a device can do scales with what it's got, and even the floor beats renting AI by the month.

Floor
A local chatbot
If a phone can only manage one thing, it's a private AI chat assistant that runs entirely on-device. No account, no monthly fee, no prompts logged to anyone's servers.
vs. a big-provider chat subscription: free, private, yours
Mid
Triage + drafting
A more capable phone classifies, summarizes, and drafts on-device around the clock at a few watts — a working assistant, not just a chat window.
vs. per-token cloud: fixed cost, nothing leaves
Ceiling
A full node
A recent phone hands heavier reasoning to a thUMBox you own on your network and acts as a true fleet node in the field.
vs. data-center dependency: your metal, end to end

The app meets each device where it is — the most a phone can run, it runs locally. Heavier work hands off to a full thUMBox you own, never to a data center. And once it's running, that same phone is how you carry your box into the field — see phones as fleet nodes.

$0 · Community plan

No box from us. No subscription. Still yours.

Bring-your-own-hardware runs on the free Community plan — the open core of the platform. You get the runtime, the dashboard, and the SDK to build and run your own packs, on your own machine, paying us nothing. This is the sovereignty path: maximum control, zero recurring cost.

Want managed first-party packs (MailBOX and the rest), idle-time intelligence, or hands-off support and backup? A paid membership adds those on top — on your hardware or ours. But it's never required to keep your box running. The free plan is genuinely free, and genuinely yours.

Install on your machine Browse the docs