← Back to Home

Flowtype

A macOS voice-to-prompt app — turning spoken, messy thoughts into structured coding prompts with one keypress

  • 5Pipeline stages
  • 0Python deps
  • 300MBLocal 4-bit model
  • Swift
  • MLX
  • Qwen3-ASR
  • LLM
  • SQLite
GitHub ↗

What it does

Flowtype is a macOS voice input app built for AI coding workflows: double-press Command to start recording, press once to output the raw transcript, or double-press to end and get an LLM-polished structured prompt injected directly into the focused field.

How I built it with AI

  • Local model, no server: transcription runs in Swift via speech-swift (Qwen3-ASR through MLX, ~300MB 4-bit), eliminating the Python server entirely — zero Python dependencies
  • Modular pipeline: Recording → ASR → PostProcess → Polish → Injection, with SessionContext state flowing through every stage
  • Product loop: SQLite (GRDB) history, stats dashboard, GitHub-style heatmap, privacy-first (audio never touches disk)

Technical highlights

  • Swift 6.2 / macOS 15+, SwiftUI
  • Qwen3-ASR local inference (MLX), AppleSpeech fallback
  • Multiple LLM providers (OpenAI-compatible), dictionary & style packs