Table of Contents

• Introduction
• Features
• Installation
• How It Works
• System Architecture
• For Developers
• Contributing
• License
• Acknowledgments

Introduction

Conversationaly is a desktop app (macOS, Windows, Linux) that captures your microphone and system audio, transcribes the meeting as it happens, and generates a summary. Transcription models and the summary LLM both run locally by default — nothing is sent anywhere. Cloud providers are available if you want them, but they are opt-in, per-feature.

It is a fork of Meetily, rebuilt around transcribe.cpp and a bundled llama.cpp sidecar. It is fully free — there is no paid tier, no license key, no telemetry.

Features

• Local transcription — 85 models across 16 families (Whisper, Parakeet, Nemotron, Canary, Voxtral, Qwen3-ASR, SenseVoice, Moonshine, GigaAM, …), downloaded on demand. Default: parakeet-tdt-0.6b-v3-q8 (1.94% WER, 25 European locales); nemotron-3.5-asr-streaming-0.6b-q8 for word-by-word live text across 32 locales.
• Live transcript — streaming-native models transcribe continuously as you speak; batch-only models are segmented by voice activity and still work live.
• Built-in AI, no Ollama required — a bundled llama-helper sidecar runs Gemma 4 locally for summaries, and can also transcribe directly as an audio LLM.
• Bring your own LLM — summaries via Built-in AI, Ollama, Claude, Groq, OpenRouter, OpenAI, or any OpenAI-compatible endpoint.
• Optional cloud STT — Deepgram, ElevenLabs, Groq, or OpenAI, if you prefer a hosted transcriber.
• Professional audio mixing — microphone and system audio captured together with RMS-based ducking and clipping prevention.
• Import & enhance Beta — transcribe existing audio files, or re-transcribe a past meeting with a different model or language.
• Summary templates — pick or write the structure your summaries follow, and set the summary language independently of the spoken one.
• GPU acceleration — Metal on Apple Silicon, CUDA (NVIDIA), Vulkan (AMD/Intel), ROCm (AMD on Linux).
• Local storage — meetings, transcripts, and models live in a SQLite database and a model directory on your disk.

Installation

Prebuilt installers (macOS .dmg, Windows .exe, Linux .deb/.rpm/.AppImage) are published on the Releases page when a version is tagged, each alongside a SHA256SUMS file to check them against.

# macOS
brew install --cask --no-quarantine bykof/tap/conversationaly

# Windows
scoop bucket add conversationaly https://github.com/bykof/conversationaly
scoop install conversationaly

The builds are not code-signed, so macOS and Windows each block them once on first launch. docs/INSTALL.md has the click-through for every platform, how to verify a download, and what to do when macOS forgets the microphone permission after an update.

Build from source

Requires Rust, Node.js, pnpm, and cmake. See docs/BUILDING.md for per-platform prerequisites.

git clone https://github.com/bykof/conversationaly
cd conversationaly/frontend
pnpm install

# macOS
./clean_build.sh

# Linux (auto-detects GPU backend)
./build-gpu.sh

# Windows
clean_build_windows.bat

Linux specifics: docs/building_in_linux.md.

Permissions

• macOS — microphone, plus screen recording for system audio (ScreenCaptureKit, macOS 13+).
• Windows — microphone; system audio uses WASAPI loopback.

How It Works

On first launch, onboarding downloads one transcription model and one Gemma 4 tier. From then on:

1. Microphone and system audio are captured, mixed, and written to a recording.
2. The same mixed audio is resampled to 16 kHz and fed to the transcription engine, which emits transcript lines as the meeting runs.
3. When you ask for a summary, the transcript goes to whichever LLM provider you configured — the local sidecar by default.

Everything above is a local process. Cloud STT and cloud summary providers are the only paths that leave your machine, and only when you select one and supply a key.

System Architecture

A single Tauri application: a Rust core (audio capture, transcription, storage, summary orchestration) and a Next.js frontend, communicating over Tauri commands and events. There is no separate server to run.

Details: docs/architecture.md.

For Developers

cd frontend
pnpm install

./clean_run.sh              # macOS: build and run (info logging)
./clean_run.sh debug        # verbose logging
clean_run_windows.bat       # Windows
./dev-gpu.sh                # Linux

pnpm run tauri:dev          # plain dev mode
pnpm run tauri:dev:metal    # force a specific GPU backend
pnpm run tauri:dev:cuda
pnpm run tauri:dev:vulkan
pnpm run tauri:dev:cpu

Architecture notes and conventions for contributors live in CLAUDE.md; GPU backend details in docs/GPU_ACCELERATION.md.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for project structure and guidelines.

License

MIT — see LICENSE.md.

Acknowledgments

• Conversationaly is a fork of Meetily by Zackriya Solutions, which it builds on under the MIT license.
• Transcription runs on transcribe.cpp, built on ggml / whisper.cpp.
• Local LLM inference uses llama.cpp via llama-cpp-2.
• We borrowed some code from Screenpipe and transcribe-rs.
• Import & Enhance was contributed by Jeremi Joslin, improved by Vishnu P S and Mohammed Safvan.
• Thanks to NVIDIA for the Parakeet and Nemotron speech models, and to the teams behind the other model families in the catalog.