freeflow-mac - GitLite

KO | EN

gitlite — search

#open-source #nlp #security #rust #ios #gemini #cross-platform #shell #python #api

freeflow-mac

(★ 9)

Private, offline AI dictation for Apple Silicon Macs using Parakeet and Qwen.

File Explorer

ZIP Download

- - ci.yml
- - freeflow-app.jpg
- Info.plist
- build_app.sh
- build_dmg.sh
- generate_icon.swift
- run.sh
- - AppConfig.json
  - CleanupPrompt.md
- AppConfig.swift
- AppStore.swift
- AudioRecorder.swift
- ContentView.swift
- DictationHomeView.swift
- DictionaryView.swift
- FreeFlowApp.swift
- GlobalHotkeyMonitor.swift
- HistoryView.swift
- LocalCleanupEngine.swift
- LocalTranscriptionEngine.swift
- MenuBarView.swift
- Models.swift
- PermissionsManager.swift
- PersistenceStore.swift
- SettingsView.swift
- TextInsertionService.swift
- PlaceholderTests.swift
.gitignore
CONTRIBUTING.md
LICENSE
Package.resolved
Package.swift
README.md
SECURITY.md
THIRD_PARTY_NOTICES.md

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

FreeFlow is a native, account-free macOS dictation app. Hold Fn, speak, and release; the app transcribes locally, cleans the transcript with a local language model, and pastes it into the focused app.

FreeFlow is open source under the MIT License. Its inference pipeline has no subscription, usage limit, telemetry, paid API, or cloud fallback.

<p align="center">
  <img src="docs/images/freeflow-app.jpg" width="860" alt="Complete FreeFlow window on macOS with Parakeet and Qwen ready for local dictation">
</p>

## Install

Download the notarized `FreeFlow-0.2.0.dmg` from the [latest GitHub release](https://github.com/rohanarun/freeflow-mac/releases/latest), open it, and drag FreeFlow into Applications.

On first launch, grant Microphone and Accessibility access when macOS asks. The application downloads approximately 1 to 1.5 GB of redistributable model assets once; dictation then works offline.

## What is implemented

- system-wide hold-to-talk with Fn and double-press hands-free lock;
- microphone capture to a short-lived local file;
- Parakeet TDT 0.6B v3 transcription through FluidAudio and Core ML;
- Qwen3.5 0.8B 4-bit transcript cleanup through MLX Swift;
- selection replacement and cursor insertion in the app where dictation started, with Accessibility-first input and a restored-clipboard Command-V fallback;
- personal pronunciation dictionary passed into the cleanup model;
- local history, raw/clean comparison, search, copy, and writing stats;
- visible microphone, Accessibility, and model readiness states;
- no login, analytics, cloud inference, subscription, or paid API.

## System requirements

- Apple Silicon Mac;
- macOS 14 or newer;
- internet access for the initial model download.

If macOS already uses Fn for its own Dictation shortcut, change that shortcut in System Settings so the two applications do not compete.

## Build from source

Development requires Xcode 26 or a compatible Swift 6.1+ toolchain.

```sh
git clone https://github.com/rohanarun/freeflow-mac.git
cd freeflow-mac
scripts/build_app.sh release
open dist/FreeFlow.app
```

Run the ordinary test suite with `swift test`. The model-backed integration tests documented in [CONTRIBUTING.md](CONTRIBUTING.md) are opt-in.

## Build a notarized DMG

The release script deliberately fails unless it can sign with a valid Developer ID Application certificate and authenticate with an existing `notarytool` keychain profile.

```sh
NOTARY_PROFILE=your-keychain-profile scripts/build_dmg.sh
```

It builds with the hardened runtime, signs the application and disk image, waits for Apple notarization, staples the ticket, runs Gatekeeper assessment, and writes a SHA-256 checksum. No signing credentials belong in this repository.

## Model configuration

Runtime choices live in `Sources/Resources/AppConfig.json`; cleanup behavior lives in `Sources/Resources/CleanupPrompt.md`. That keeps model selection and correction policy editable without baking behavior into the Swift pipeline.

For an English-only build, change `asrVersion` from `v3` to `v2`. FluidAudio documents v2 as the higher-recall English option; v3 covers 25 European languages.

## Shipping models inside the app

The default build downloads models on first launch because this keeps the app bundle small. Both chosen model licenses permit redistribution with their required notices. For a fully preloaded installer, stage FluidAudio's compiled Core ML directory and the MLX model directory, include them as app resources, and initialize each engine from those local directories. Keep the attribution file in the distributed app.

## Scope

This is an original implementation of the local dictation workflow described on the referenced product page. It does not copy Rhino Voice code, branding, or assets. It is not a meeting recorder or podcast editor.

## License

FreeFlow's original source is available under the [MIT License](LICENSE). Its dependencies and downloadable model weights retain their respective licenses and attribution requirements; see [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).