Something clicked when I gave a small local model a tool.
"List files in this folder, rename anything that matches this pattern."
It worked. Fully offline. On my M1 Pro with 16GB RAM. No cloud, no API key, no subscription. Just a 9-billion parameter model doing something useful on my machine.
That's when I started building Familiar.
Familiar is an open source macOS app (iOS companion coming) that runs a local AI agent on your device. Not a chat interface. An actual agent that can take actions. Here's where it is after the first week:
Hardware detection on first launch. The app looks at your RAM, chip, and available resources and recommends the best model you can run without slowing your machine down. You don't have to research this yourself.
A default small model ships with the app. You're running a local agent in minutes, not after a 20-minute download.
File tools out of the box. Create, rename, move, delete. Basic but real. Works without internet, without API credits, without a monthly charge.
Night shift mode (in progress). When you close the lid and go to sleep, the app switches to a larger, more capable model that uses your full available resources. You wake up to work done by a smarter local agent. This is the feature I'm building toward most aggressively.
The key insight from testing Qwen 3.5 9B: small models don't need to be smart about everything. They need to know what tool to pick next. That's a much more achievable bar. And tool calling support has matured fast — Qwen 3, Phi-4, and Mistral all support it natively now.
Is it as capable as Claude Opus 4.6 or GPT-5.4? No. For private local tasks, offline work, file management, and anything where data shouldn't leave the machine? It's already good enough, and it's getting better.
Familiar is open source. GitHub going public soon. iOS companion is in progress.
Full build log: https://thoughts.jock.pl/p/familiar-local-ai-agent-mac
Newsletter on AI agents and practical automation: https://thoughts.jock.pl
Top comments (0)