Bring your own keys

What you supply

The configuration screen asks for a provider, a model, the API key for that provider, and a Telegram bot token. As of 2026-09-05 the providers offered are Anthropic (Claude Opus 4.7), OpenAI (GPT-5.5), and Google Gemini (Gemini 3.1 Pro) — one model each. The screen has no server, SSH or network fields.

Your key, your limits

Because the key is yours, four things follow, and none of them is a policy we chose:

• The model bill is yours. Your subscription pays for the box. The model calls are made with your key against your provider account, so they are billed there, on your provider's rates.
• The rate limits are yours. Whatever tier your provider account sits on is the ceiling the agent works under. If your account is throttled, the agent is throttled.
• The outages are yours. When your provider is down, the agent has no model to call. The box is still running; the model behind it is not answering.
• Revocation is yours. You can revoke the key in your provider account at any time, without asking us. The agent stops being able to call the model as soon as your provider stops honouring the key.

What we do with the key

The key is sent to the box that is claimed for you, so that the agent on it can call your provider. Two things are true of the API that carries it: it masks the key in its own server-side log line, and it does not return the key, or any other value you typed on that screen, in a response.

The Telegram bot token you supply is the one your box talks to you through. The same API does not write it to that log line.

Choosing a key to give it

Use a key you can revoke on its own. Provider accounts differ, but where your provider supports more than one key, a key created for this box and used by nothing else is the one to paste, because revoking it later costs you nothing else.