Do I need to know Docker or CUDA to self-host an AI agent?

Not with this one. The runtimes, the drivers and the weights are installed and tested before it is boxed, so there is no compose file to write and no quantisation rabbit hole to go down. The parts that usually eat a weekend are the parts that are already done.

Is a self-hosted agent cheaper than a subscription?

It depends how long you keep it. The hardware is €549 once and there is nothing to pay us again, so it stops being a monthly line item — but you still pay your model provider for tokens, exactly as you do now. What you stop paying is the markup a hosted service adds on top of them.

Can I still use Claude or GPT if I self-host?

Yes, and that is the intended way to run it. Paste in the key you already pay for and the box calls the provider directly. Self-hosting moves the agent, not the model — you keep the frontier models and just stop renting the loop around them.

A self-hosted AI agent, and the machine it needs

A self-hosted AI agent is one that runs on hardware you control rather than in somebody else's account. The model can still be Claude or GPT — what changes is where the loop runs, where the keys sit, and who bills you.

The agent is not the model

The two get conflated constantly, and the distinction is the entire reason a box like this exists. The model is the thing that produces text — Claude, GPT, Gemini, or an open one like Llama or Qwen. The agent is the loop around it: the part that holds the task, calls the model, runs a tool, reads the result, and decides whether it is done.

That loop is cheap. It is a small program that spends most of its life waiting. What it is not is something a browser tab can do while it is closed — and that is where a self-hosted agent stops being a preference and starts being the only arrangement that works.

Why it has to stay awake

Almost every agent worth running is a long job. Research that takes forty minutes. An inbox triaged before you get to it. A scheduled review that fires at 6am. A laptop is the wrong host for all of these, because a laptop's normal state is shut.

A dedicated machine solves it by being boring: HermesBox draws about 20 W, roughly a fifth of a lightbulb, and sits on your network doing the work while everything else in the house is asleep. It messages you when something needs you rather than waiting to be asked.

Bring your own key, and keep your own invoice

The commercial difference between self-hosting and subscribing is who holds the API key. A hosted agent service holds it on their servers, resells you tokens with a margin on top, and bills you in their own credits — so you pay twice for the same model and can be repriced at any time.

Self-hosted, the key stays on your hardware and calls the provider directly. Anthropic, OpenAI, Google or OpenRouter bill you at their rates, on the account you already have. There is no proxy in between and no cut taken. The €549 is the whole commercial relationship.

What you actually have to do

The honest answer is: less than the phrase "self-hosted" suggests, provided the hard part is done before it ships. HermesBox arrives assembled, with the runtimes installed and the local weights already on the 512 GB drive, so the first boot is a first boot rather than an install.

Power and Ethernet, paste in the keys you already have, and message it from Telegram or the web UI. If you want to go deeper — different models, a different context length, reaching it from outside the house — all of that is there and none of it is in the way.

Common questions

Do I need to know Docker or CUDA to self-host an AI agent?: Not with this one. The runtimes, the drivers and the weights are installed and tested before it is boxed, so there is no compose file to write and no quantisation rabbit hole to go down. The parts that usually eat a weekend are the parts that are already done.
Is a self-hosted agent cheaper than a subscription?: It depends how long you keep it. The hardware is €549 once and there is nothing to pay us again, so it stops being a monthly line item — but you still pay your model provider for tokens, exactly as you do now. What you stop paying is the markup a hosted service adds on top of them.
Can I still use Claude or GPT if I self-host?: Yes, and that is the intended way to run it. Paste in the key you already pay for and the box calls the provider directly. Self-hosting moves the agent, not the model — you keep the frontier models and just stop renting the loop around them.