Models

Jak uses model API access that you configure. Your model provider bills usage separately from Jak. Use the exact API model ID supplied by that provider.

Add and test a model

1. Open Models → Add model.
2. Choose Provider and enter Model ID.
3. Use the prefilled Endpoint for a built-in provider. For a compatible endpoint, enter its API URL.
4. Enter the API key, or reuse a saved key for the same provider and endpoint when offered.
5. Select Use as installation default if this should be the team’s fallback.
6. Choose Save model, then Test capabilities on the model’s card.

The test reports Text generation and Tool calling separately. A failed text test prevents work from starting. A model without working tool calling cannot use connected research tools.

Choose which model does the work

The installation default is the fallback. An agent’s Team → Model override can choose a different saved model.

In chat, the top-bar Model selector can choose a model for chat and tasks created from it. In New task, Model for this task can choose one model for all assigned agents. A task’s explicit model selection also applies to retries.

Choose Use each agent’s model setting when specialists should keep their individual settings. A saved task can therefore involve different models across its team.

Use a compatible or local endpoint

The app includes built-in endpoint options for OpenAI, Anthropic, Qwen, and DeepSeek. Other endpoints, including local models, must be allowed by the host through JAK_MODEL_ENDPOINTS before saving or testing them.

Ask the person running your Jak installation to allow the endpoint. The browser form alone cannot change that host setting. Keep the API key specific to the service behind that URL.

Troubleshoot model access

If text generation fails, check the exact model ID, endpoint, API key, provider access, and available billing balance. Read the diagnostic shown after the capability test.

If text works but tools fail, choose a model and endpoint that support tool calling and test again before relying on account data. Without a configured model, chat may show an offline preview; it is not a completed model-generated task.

Credentials are encrypted on the Jak server. They are still part of your installation’s protected data, so keep its host and backups private.