Skip to main content
Devic is model-agnostic: an agent names a model, and Devic calls whichever provider serves it with your credentials. Providers are configured once for the account. Model providers

​
Supported out of the box

OpenAI

Anthropic

Google Gemini

xAI

DeepSeek

Kimi

Google can be reached either through the Gemini API with a key, or through Vertex AI with a service account — the same models, billed and governed through your Google Cloud project.

​
Credentials

Each provider takes its own key. Keys are encrypted at rest and masked on read, and can be held in your own vault instead, so the value never lives in Devic at all.
Test the credential when you save it. A provider that is configured but whose key is wrong shows up as a failed run much later, in whatever agent happens to use it first.

​
Custom providers

Anything with an OpenAI-compatible API can be added as a custom provider: a self-hosted model, a regional gateway, an internal inference service. You supply the base URL, the credential and the model names, and it appears in the model selector alongside the rest.
Custom providers are as good as their compatibility. If the endpoint does not return tool calls in the OpenAI shape, agents using tools on it will not work — an assistant that only converses will.

​
Choosing a model per entity

The provider is account-wide; the model is chosen per agent and per assistant. That is what lets a cheap model handle classification while an expensive one handles the reasoning, on the same account and the same key. Cost is recorded per model, per agent and per tenant, so the trade-off is measurable rather than assumed. See Costs.

Secrets and vaults

Keeping provider keys outside Devic.

Costs

What each model is actually costing you.