AI agent hostingnot on your computer.

The agent stops calling the model and tells you in the chat. It stays online, keeps its files and memory, and everything resumes the moment you top up. Nothing is deleted and no charge appears without you choosing it. There is also a daily cap, so a runaway loop can burn at most one day's worth before it stops itself.

Your messages and the agent's answers go through our own proxy on our German node to the provider of the model you picked, routed via OpenRouter, with the model provider (Anthropic, DeepSeek, OpenAI or Nous Research) behind it, which may be outside the EU. Our proxy stores only metadata for billing: time, model, token counts and cost, not the content of your messages. The conversation history, memory and files live inside your own container in the German data centre and are deleted with the server. We can technically access that container like any hosting provider, but we only do so for support you asked for or a substantiated abuse report. The privacy policy has the full detail.

You can. It is MIT licensed and the install is documented. Three things usually change people's minds. First, it is only awake while your machine is, so no scheduled jobs and no answering Discord overnight. Second, an agent on your desktop has your file system and your home network, and it acts on text it reads from web pages and messages, which is a genuine attack surface. Third, you become the operator: updates, Python environments, crashes at 2 a.m., API keys in a local file. We take all three.

You create a bot in the Discord developer portal, paste the token into your agent's settings, and invite it to your server. From then on it reads and replies in the channels you allow. You decide whether it answers everyone or only specific user IDs, and by default it only listens to you. Note that message content then also passes through Discord under Discord's own privacy policy.

Not at the moment, and it is a deliberate choice. The spending cap, the daily limit, the rate limiting and the platform rules are all enforced by our inference proxy, which sits between the agent and the model provider. With your own key the agent would call the provider directly and none of those protections would exist, including the one that stops a hijacked agent from spending your whole account balance in an afternoon.

It depends entirely on the model, which is why the panel shows a factor next to each one. On the cheapest model, €10 of credit is roughly fifteen thousand ordinary messages. On Claude Sonnet 4.6 it is a few hundred. Most people run everyday chat on a cheap model and switch to a strong one for the handful of tasks that need it, which is what the model switcher is for.

No. Agents are firewalled off from every other container, including your own game servers, and from our host and panel. If you want your agent to touch another server, you connect it deliberately through a documented interface, never by it simply sitting on the same network.

Those are chat windows you open. This is a process that keeps running: it has its own terminal, its own files and its own memory, it can be triggered on a schedule, and it lives in your Discord or Telegram rather than in a browser tab. You also pick the model per task, including Claude Sonnet, and you see exactly what each answer cost.

Monthly, from the panel, effective at the end of the paid period. No contract, no minimum term, no cancellation fee. Remaining credit stays usable until the period ends.

An open-source AI agent by Nous Research, released under the MIT licence. It runs a loop of thinking and using tools, so it searches the web, runs commands, reads and writes files and remembers things, instead of only answering in a chat window. Renzom hosts and operates it for you. We are not affiliated with Nous Research.