The closed loop
AI server.

Open OverwatchOS from a browser on your network. It asks what you would like to do today and offers suggestions to start from. Describe the helper you want in plain English, such as answering questions from a folder of contracts or drafting replies that wait for your approval, and OverwatchOS builds it with guardrails you set: what it can read, what it can send and whether it can reach the internet at all. Models, memory and documents stay on the box. Any OpenAI compatible tool can point at the Node by changing one URL.

Quantized open weight language models, vision, speech, embeddings, agents and RAG. The sweet spot is 7B to 35B parameter models. A dense 70B model loads in 64 GB but runs at 1 to 2 tokens per second, so treat it as an occasional tool rather than a daily driver. Frontier cloud models remain larger than anything this box can run, which is what Connected Mode is for.

Local Mode runs models, memory and RAG on the Node with no cloud AI, and lets you keep integrations you trust switched on, such as a webhook to your own server. Closed Loop Mode blocks every outbound connection at the gateway and the host firewall for the session, so nothing leaves the box. Connected Mode adds cloud models you approve, using your own keys, with a log of what left and why.

No subscription, no meter, no account with us. Core operation is local. Internet is needed only when you choose it: downloading new models, installing software updates, or using Connected Mode.

The components are open source today: Ubuntu LTS, llama.cpp, Ollama, Whisper, the vector database and the workflow engine. The OverwatchOS layer that ties them together is published with its license alongside the first shipping image. Nothing on the Node stops working if we do.

Yes. The Node exposes an OpenAI compatible API on your network. Change the base URL and go. In Local and Closed Loop modes, cloud model names simply do not resolve, so a misconfigured tool fails loudly instead of leaking quietly.

Production begins when all 100 first run slots are reserved. Delivery is estimated at 11 to 13 weeks after production begins, with free shipping. Your payment is held in escrow by Dough, the preorder platform, until the run funds. If the run does not fund or the product cannot be delivered, the reservation is refunded under Dough's terms.

Overwatch Node designs the product and the OverwatchOS image. Dough hosts the storefront, checkout and escrow and coordinates production with a contract manufacturer; Dough does not build hardware itself. Each unit is imaged and burn tested before it ships. Warranty and support terms are published on this site before production begins, and every reservation holder is notified before they take effect. Until then your money is protected by escrow, and you can email karl@quotabuster.net with any question.

You can, and it is a fair comparison because a DIY HX 370 mini PC uses the same silicon. Ollama plus Open WebUI gets you private chat and document questions in an afternoon. What you assemble yourself over the following weekends is the rest: the network policy that decides what may leave, the agent runtime, persistent memory, the audit log and the maintenance. The compare page puts the costs side by side.