A new open-source macOS app called Familiar runs a local AI agent entirely on your device - no cloud connection, no API key, no subscription fees.
The project started from a specific observation: a small language model with decent tool-calling ability (meaning it can decide which action to take next, like renaming a file or moving a folder) is already useful for real tasks. Developer Pawel Jozefiak tested Qwen 3.5 9B on an M1 Pro MacBook with 16GB RAM and found it could reliably handle file operations completely offline.
What It Does Right Now
Familiar detects your hardware on first launch and recommends the best model your machine can run. It ships with a compact default model so you can start immediately without downloading anything large.
Current tools are focused on file management: create, rename, move, and delete files, all processed locally. The most interesting feature is "Night Shift" - when you close your laptop lid, the app automatically switches to a larger, more capable model that takes advantage of freed-up RAM and compute while you sleep.
The developer is building this on deliberately mid-range hardware (M1 Pro, 16GB) rather than optimizing for high-end machines, which should mean it works acceptably on most Apple Silicon Macs from the last few years.
Where It's Headed
Planned additions include web access, clipboard utilities, shell commands, and an iOS companion app. The GitHub repo is expected to go public soon.
The local-first approach trades capability for privacy and cost. You won't get GPT-4 or Claude-level reasoning from a 9-billion-parameter model running on a laptop. But for straightforward file automation tasks where you'd rather not send data to a server or pay per API call, the tradeoff makes sense. The Night Shift feature - offloading heavier work to idle hours with a bigger model - is a clever way to squeeze more capability out of limited hardware without interrupting your daytime workflow.