Atako inference

Every agent runs on a large language model. When you create an agent, you choose where the model calls go:

​

How Atako inference works

Atako inference is powered by OpenRouter, which gives access to hundreds of models from many providers (Anthropic, OpenAI, Google, Mistral, and others). When your agent calls a model, the request goes through OpenRouter to a provider that serves that model. A model can be served by several providers, and OpenRouter picks one for each request.

​

Choosing a model

When you create an agent, you pick its model from a catalogue of the models available through OpenRouter:

• Search the whole catalogue.
• Favorites — star the models your team uses most. Favorites are shared across your company.
• Already used — the models your company has already launched agents on.
• Suggested by Atako — a short list of models Atako recommends.

If the full catalogue can’t be reached, the picker falls back to Atako’s suggestions. Some models listed by OpenRouter aren’t offered on Atako inference; the picker tells you when that’s the case, and you can still use them with your own OpenRouter key.

​

Each provider has its own terms

The provider serving a model sets the terms of use, not Atako. Terms differ from one provider to another, including on data retention and on whether the data you send may be used to train models. The prompts, files, and tool results your agent sends to the model reach that provider. Atako is not responsible for how a provider uses that data. The provider information published by OpenRouter may be incomplete or out of date. Checking the terms that apply to your use is your responsibility.

​

Need contractual guarantees?

Run your agents on your own provider key. Your existing contract with the provider then governs how your data is handled.