The basics, and how it works under the hood.

No. Once your access opens, you can just use the chat. Adding a Mac grows the network, and your Mac earns MAGIC when it answers requests.

During the beta, your prompts are processed on computers in the Feya network, including Macs and PCs run by other Feya members, which can see what you send while they answer it. This is a concession we are making for the beta, so please don’t share anything sensitive. Feya does not store your chat history during the beta: your conversations are not saved on our servers, in your browser, or on the computers that answer them (except that, rarely, part of a conversation can end up in an error log on our servers or on those computers, as the full notice explains).

Privacy notice

AI requests are free during the beta. We’re still finalizing the full program terms, and will publish them before they take effect.

To pause, flip the switch in the menu bar. To remove it, quit Feya from the menu, drag it from Applications to the Trash, and delete the .feya folder in your home folder, where the models live, and the FeyaAI folders in Library/Application Support and Library/Logs.

Tap Sign in and enter your email. We’ll send a sign-in link that works for 15 minutes.

The chat sends it to Feya’s coordinator. The coordinator tracks which devices are online, which model each has loaded and how fast it is, and routes the request to a device, or a group of devices, running that model. The answer comes back the same way.

Two widely used engines: llama.cpp for text models and stable-diffusion.cpp for images, running on your Mac’s GPU through Metal. The models are open-weight: Llama, Gemma, Qwen, DeepSeek. Feya’s own code just connects these engines to the network.

The coordinator picks the largest model the online Macs can hold between them. One Mac leads, the others help, and the model’s layers are split by how much memory each one has. The lead runs the model; the helpers compute their layers over llama.cpp’s RPC protocol. It works best when the Macs are close to each other on the network.

Your Mac opens an outbound WebSocket to Feya, encrypted with TLS, and keeps it open. Requests come in over that connection, so nothing on the internet connects to your Mac directly. Each device has its own key and signs its connection to Feya with it.

You don’t have to choose. Your Mac reports how much memory it can spare (part of its RAM, capped by what the GPU can handle), and the coordinator gives it a model that fits, or makes it a helper on a bigger one. A small starter model downloads first, and a better one follows in the background.

Each answer produces a receipt: model, tokens and time, signed with an Ed25519 key that never leaves your device. The coordinator checks the signature and credits MAGIC based on the tokens, the model file’s precision, your Mac’s speed and your device’s reliability rating. When a model is split across several Macs, the Mac that coordinates the split gets 5% of the request’s reward and the other 95% is divided by how much of the model each Mac held. A Mac we can’t confirm took part isn’t paid until we can.