Model selection

The model selector#

When you create or configure a desktop, the model selector lets you pick which AI model your agent will use. It is a searchable dropdown that fetches available models directly from your provider's API.

Start typing a model name ("claude", "gpt", "llama") and the selector filters results in real time. Pick one and it is saved to your desktop configuration.

How it works#

The selector queries your provider's API using the key you entered:

• Anthropic lists Claude models available on your account.
• OpenAI lists GPT and other models available on your account.
• OpenRouter lists 200+ models from dozens of providers (Anthropic, OpenAI, Meta, Mistral, Google, and others).

The list reflects what your API key actually has access to, so you only see models you can use.

Choosing a model#

Different models have different strengths:

Coding agents#

Claude models are generally the strongest for autonomous coding. They handle long contexts well, follow instructions precisely, and produce clean code. Claude Sonnet is a good balance of speed and capability for most coding tasks.

General purpose#

GPT-4o is a solid all-rounder. If you need broad task coverage without optimizing for a specific domain, it is a reliable default.

Budget-conscious use#

Smaller models like Claude Haiku and GPT-4o-mini cost a fraction of their larger counterparts and still handle straightforward tasks well. OpenRouter also gives you access to open-source models that are significantly cheaper.

OpenRouter: the multi-model option#

If you are not sure which provider to go with, OpenRouter is worth a look. A single API key gives you access to models from every major provider plus dozens of open-source options.

This is useful for trying different models without managing multiple API keys, switching to cheaper models for simple tasks, and accessing open-source models like Llama, Mistral, or Qwen.

Changing your model#

You can change your agent's model at any time from the desktop settings. The new model takes effect on the next conversation. No restart needed.