A connection describes one LLM endpoint: its provider, the API key and whatever else that provider needs. Connections are named and there can be as many of them as you like – one per provider, several accounts with the same provider, or a cloud provider next to a locally hosted model. Agents then refer to connections by name, so you can point different agents at different providers.

The table lists the connection Name, its Provider and the Endpoint it points at (empty for providers that have a single fixed endpoint). The buttons on the right manage the list. Removing a connection also deletes its API key from the secure storage, once you confirm the page with OK or Apply.

Each connection is edited in its own dialog. Enter a Name – it is what you will pick from in the agents’ grid, so name connections after what they are for – choose a Provider, and fill in the fields that provider requires. Every field carries an info icon with a description of what belongs in it.

The following providers are available:

The Test connection button verifies the connection: it loads the list of models the endpoint offers and sends a short request to check the credentials. The loaded models are remembered and offered in the agents’ grid afterwards.

Different agents can use different models – this approach allows you to optimize agent capabilities and its cost (frontier models are more expensive, but not every task requires them).

AI Assistant uses two types of agents – reasoning agents and task agents – which serve different purposes. Reasoning agents are the ones running larger and more complex tasks such as work planning, requirements analysis and so on. Task agents are typically smaller and perform more limited tasks such as documentation search, data analysis, etc.

The basic configuration allows you to configure connection and exact model for each agent type as well as fine-tune advanced model properties:

A good approach is to select more powerful models for reasoning agents and select smaller (and usually faster and cheaper) models for task agents. You can even use local models for task agents to significantly reduce the cost of running AI Assistant.

You can configure models on a per-agent basis if you’d like by opening detailed agent configuration with Per-agent models (advanced) button.

The configuration for each agent works in the same way as for agent types above – select a provider connection and pick a model for each agent. To learn more about each agent and their usage, see Agent types.

To make configurations easier to manage, you can create profiles. A profile is a named set of choices – which connection and which model each agent uses, plus optional per-agent model properties. You can keep several profiles side by side, for example a fast and cheap one for everyday work and a frontier-model one for hard design tasks, and switch between them.

You can use Export button below these settings to export the configuration as json file. This can be very useful if you need to copy the configuration to another machine or as a simple backup. Use Import button to import the json file.

Use the Test configuration button to verify your settings. Assistant will connect with every provider used in current set-up and will test each model as configured for different agents. This is the quickest way of verifying that the configuration you have is valid.

Advanced configuration allows you to configure additional settings that tune the behavior of Assistant, its token consumption, logging details or even zoom level for Assistant view.

The configuration uses few basic terms: