# Rift — agents.md

> Machine-readable summary for AI assistants, search, and tooling.  
> Site: https://myrift.dev · Source: https://github.com/mkmk10k/rift

## What it is

**Rift** is a macOS app (Apple Silicon, MLX) for **voice-to-text** and **text-to-voice**, entirely on-device. No cloud for speech or transcripts. Optional network only for first-run model downloads and optional “Check for updates.”

## Who built it & why

Built by **Mikko Kiiskila** out of personal need: **XLRS** (low vision), **dyslexia**, and **ADHD**. The homepage explains how product decisions map to those experiences (patience, first-word capture, natural TTS, silence polish, etc.).

## Core stack

| Layer | Tech |
|-------|------|
| ML framework | **MLX** (Apple Silicon) |
| Speech-to-text | **Parakeet** (MLX) |
| Text-to-speech | **Kokoro**, optional **Chatterbox** variants |
| Text intelligence | **Gemma 4 + Qwen3** local LLMs — merge, correct, extract, polish, **Code Talk** (developer TTS transform) |

## Notable features

- **No auto-endpointing** — you stop when ready; no 2s silence cutoff.
- **250ms lead-in buffer** — first syllable preserved.
- **Live paste** — streaming text into the active app.
- **Silence polish** — after ~5s pause while dictating, on-device polish can run.
- **Polish modes** — verbatim / clean / professional.
- **Code Talk** — in IDEs/terminals/docs, transform technical text before TTS.
- **Global shortcuts** — ⌃1 read, ⌃2 dictate, ⌃3 show/hide/pause widget.
- **Audio cues** — subtle tones for record start/stop.

## Requirements

- macOS 14+
- Apple Silicon (M1+)
- ~2GB models on first run

## License & download

- **MIT** open source  
- Releases: `https://github.com/mkmk10k/rift/releases/latest`

## Contact

- Issues: GitHub  
- Community: see site footer (email placeholder may be updated when live)