Your model.
Your call.
Choose where the intelligence runs, and change it as your work changes.
Cloud providers
Asyncat's catalog lists more than 35 providers. Among them are OpenAI, Anthropic, Gemini, xAI, Mistral AI, DeepSeek, Alibaba Qwen, Kimi, Groq, Together AI, Perplexity, Cohere, Fireworks AI, Cerebras, DeepInfra, Hugging Face, NVIDIA NIM, OpenRouter, Amazon Bedrock and Azure OpenAI. Custom connects any other OpenAI-compatible service.
Use your own API key and a model your account can use. For example, the catalog names OpenAI's GPT-6 models (Astra, Sol and Luna) and Anthropic's Claude 5 family, such as claude-opus-5-5. Provider model names, prices and limits change; the provider's own list is the source of truth. Requests to a cloud model are handled under that provider's terms.
Local models
Asyncat has a built-in llama.cpp engine for GGUF models. Under Customize → Models → Language → Local models, download a model or load one you already have; models in your Hugging Face cache are listed too. The engine is installed and updated from the same page.
| Runtime | How it fits |
|---|---|
| Built-in llama.cpp | Runs GGUF models that Asyncat manages for you. |
| MLX | Runs Hugging Face MLX models on Apple Silicon, or on Linux, through mlx_lm.server. |
| Ollama, LM Studio | Connect to the local server they run, at the address they show. |
| Other local servers | vLLM, Jan, LocalAI, KoboldCpp and GPT4All are in the catalog, and Custom takes any other OpenAI-compatible address. |
A local model loads when a chat that uses it sends a message, and the last model you loaded loads again when Asyncat starts. The built-in engine holds one model at a time, so a chat on another local model waits until the current one is done. Once a model is downloaded, it works without an internet connection. Tools that browse the web or call a service still need the network.
Getting connected
- Open Customize → Models → Language and choose Add provider.
- Search for the provider and paste your API key. For a local server, enter the address the runtime shows.
- Pick a model from the list Asyncat fetches. If the list doesn't show the one you want, type its Model id. Anthropic doesn't share its list this way, so type the model ID there.
- Send a short message to check the connection.
Keys are stored encrypted, with the key in your system keychain. When you report a connection problem, remove keys and other credentials from screenshots and logs. See Troubleshooting for local engine issues.
A model per chat
The model menu under the message box picks the model for the chat you're in, so two chats can use different models at once. Picking one changes that chat only, and it becomes the default for new chats. Models shows it as the default model for new chats.
- A chat remembers its model from its first message, so changing the default later doesn't move it.
- Cats and Quick ask use the default model, unless a Cat has a model of its own.
- An agent with a model of its own answers with that model, whatever the chat uses.
See Chats for the rest of what you pick under the message box.
Choosing a model
Models differ in how well they use tools. One that chats well may be less reliable over a long run of actions, so test a simple task before handing it a big one. With Tools set to Auto, a chat uses agent tools when a project is open or the model supports them.
To keep prompts small for local models, each agent step sends a core set of tools rather than the full list: about 40, or 24 for models that read tools from the prompt. When the agent needs another tool, it finds it with tool_search. For models that reason, the + menu sets the reasoning effort.
Voice, images and training
The Models page also covers Voice, Images, Usage (requests and tokens per model) and Training. See Voice & images and Training.