Models keep getting smarter in labs. For most people, daily life barely changes. Friday is my attempt to close that gap: an assistant that lives on your Mac, sees what you see, and talks back.
Press a hotkey, or say "Hey Friday", and ask about what is on your screen. Friday captures the display, reads it with a local vision model, thinks with a local language model, and answers out loud. No cloud. No API keys. No subscription.
The core loop works today: speech-to-text, screen context, local inference, and text-to-speech in under a couple of seconds. Friday remembers you across sessions, manages its own model servers, and can take action through GitHub, Calendar, Gmail, Notion, and a browser when you ask it to.
I'm building Friday because I want the thing on my desk to feel as capable as the models in the data centre without sending my screen or my life to someone else's server. It is a consumer product, not a research demo.
V1-V3 shipped (June 2026): Voice loop, memory, self-managed runtime, agent tools, wake word, and early browser control. Mac control and proactive briefings are next.
Tech stack
| Layer | Technology |
|---|---|
| App | Swift / SwiftUI (macOS menu bar) |
| Brain | Qwen3-8B Q4 via mlx-lm (:8080) |
| Screen | FastVLM-0.5B via mlx-vlm (:8082) + Vision OCR + accessibility tree |
| Voice in | Apple Speech (on-device) + Parakeet optional |
| Voice out | Kokoro TTS (:8081) |
| Agents | Tool-calling loop: GitHub, Notion, Gmail, Calendar, Playwright browser |
| Privacy | Runtime-local inference; connector auth optional via OAuth or Composio |