Documentation
Getting Started
From download to your first local conversation in a few minutes. needs a subscription to unlock, and from then on it runs on your own machine -- nothing leaves your device unless you ask it to.
Requirements
macOS 11+ on Apple Silicon (M1 or newer) and Linux (x86_64 or aarch64, glibc 2.39 or newer) -- available now. Windows is queued.
Optional. Windows CUDA acceleration needs an NVIDIA driver from the R580 branch or newer; the newest CUDA features need R615 (615.xx or newer). Without it, RunMyAI runs on Vulkan or the CPU.
Room for the models you download -- a few GB each, your choice.
A subscription unlocks the app. Subscribe on this site, then activate this machine with the token from your portal. Models, chats and files stay local.
Needed to download models, web research, email, and phone channels. Everything else works offline.
Five steps
Install and activate
Download the build for your platform and open it. It ships with an Apple Developer ID signature and notarization, so it launches without extra runtimes or installers. The app then asks for the activation token from your portal, which links this machine to your subscription.
Pick a model for your hardware
RunMyAI detects your CPU, RAM, and GPU and recommends models that fit. Browse over 2 million open models from HuggingFace (GGUF + MLX), and check the ModelIndex to see what runs best on machines like yours. Already running Ollama, LM Studio, Jan or another local runtime? Its tab imports the models you already have, without downloading them again.
Your first conversation
Open MyChat and start typing. Inference runs entirely on your device -- zero API cost, nothing leaves your machine. Paste an image for vision, drop a document for RAG, or speak for voice.
Let it operate
As you work, the Intelligence Nucleus plans across domains. Set autonomy per lane, approve sensitive actions, and connect a phone channel like Telegram or iMessage to command it from anywhere.
Own your data
Everything lives in a local SQLite database you can browse, query, and export with MyData -- and back up or restore at any time with MyBackup. It is your data, wide open.
Model formats: GGUF and MLX
The universal format. Runs on any supported computer through the built-in llama.cpp engine, and handles several tasks at the same time.
Apple Silicon's native format, run by the built-in MLX engine. Often faster on a Mac, and runs several tasks at once, vision models included.
There is no engine to pick. When you load a model, RunMyAI reads its format and starts the matching engine. The GGUF or MLX badge next to each model shows which one runs it.
MLX needs a Mac with an M-series chip (M1 or newer). On any other machine MLX models do not load -- choose a GGUF build of the same model instead.
When you search HuggingFace, the GGUF and MLX boxes choose which formats the results include. The MLX box is available only on Apple Silicon; hover it to see why when it is off.
Tick "Only include models that can fit on my device" to hide models too large for your memory. The largest size your machine can run is shown on the right of the filter bar.
is available now for macOS (Apple Silicon). A subscription unlocks the app on your machine.