Records mic and system audio, transcribes live with Whisper or Parakeet, summarizes with the LLM you pick. Nothing leaves the machine unless you choose a cloud model.
Fork of Zackriya-Solutions/meetily. Tracks upstream main, adds a live in-meeting assistant.
- Download (recommended): macOS Apple Silicon .dmg or Linux x86_64 .deb, checksums in Releases.
- Build from source: Windows, Intel Mac, CUDA. See below.
The macOS app is ad-hoc signed, not notarized. After dragging it to Applications, run xattr -dr com.apple.quarantine "/Applications/Meetily Pro.app" once. It needs microphone and screen-recording permission (that is how system audio is captured).
- First launch downloads a transcription model. Whisper
baseis fine to start. - Pick an AI model: built-in local model, Ollama, a cloud key (OpenAI, Claude, Groq, OpenRouter, custom OpenAI-compatible), or your own Claude Code or Codex sign-in.
- Choose mic and system-audio device, press Record. Stop when done; the summary appears in the meeting.
- When a meeting ends, the assistant posts a short summary right into the chat; just keep asking if you want more detail.
- Summaries are shorter: what the meeting was about, what was decided and what happens next.
- The chat memory gauge now sits next to the message box and turns into a Compact link when the chat gets long.
- Copy any assistant reply with one click.
- Paste long texts into the assistant chat; a gauge shows how full its memory is, and Compact keeps a short summary so you can keep going.
- Live tips follow what you ask in the chat and the whole meeting so far, and arrive one to three at a time.
- Questions and tips from a call stay on the meeting page until you close them.
- Pick a model with the Use button next to each ready AI service.
- Use your Claude Code or Codex sign-in as the AI model, no API key needed.
- The live assistant now brings useful facts and tips, not only questions.
- Change the transcription language during a call (Whisper).
Live transcription on device (Whisper with Metal/CUDA/Vulkan/ROCm, or Parakeet on CPU), VAD so only speech hits the model.
Live assistant beside the transcript: questions and tips cited to transcript moments, plus a chat about the call so far. Design: docs/LIVE_ASSISTANT_DESIGN.md.
When a call ends, the assistant posts a short recap into the meeting chat, and the chat stays open for follow-up questions about the meeting.
Import an audio file, or re-transcribe a meeting with another model or language.
Mic and system audio mixed with ducking and clip protection; everything stored as files plus one SQLite database.
- Otter, Fireflies, Granola: better diarization and integrations, cloud processing and storage.
- MacWhisper, raw whisper.cpp: great local transcription, no system-audio capture, mixing or summaries.
- Meetily Pro: the whole pipeline, local by default, model is your choice. No bot, no diarization yet.
All settings are in the app. Data lives in the Tauri app data dir (~/Library/Application Support/com.meetily.ai, %APPDATA%\com.meetily.ai, ~/.local/share/com.meetily.ai): meeting_minutes.sqlite, models/, and a recordings folder with one directory per meeting.
Rust stable, Node 20, pnpm 9.15.9, CMake. Linux adds clang and libasound2-dev; Windows adds VS Build Tools C++ workload.
git clone https://github.com/asidko/meetily-pro
cd meetily-pro/frontend
pnpm install --frozen-lockfile
./dev-gpu.sh # dev run, auto-detects GPU, builds the llama sidecar
./build-gpu.sh # release bundle in ../target/release/bundle/Windows: .\dev-gpu.ps1 / .\build-gpu.ps1. Force a backend with TAURI_GPU_FEATURE=cuda|vulkan|hipblas|metal|coreml|openblas|none. Details: docs/BUILDING.md, docs/GPU_ACCELERATION.md.
Tests: pnpm test in frontend/ (one Bun process per file), cargo test -p meetily at the root. Lint: pnpm lint.
One Tauri 2 process. Rust owns audio capture, mixing and VAD, transcription (whisper-rs, ONNX Runtime), SQLite, and every LLM call. The UI is a static Next.js export that talks to Rust only via Tauri commands and events; its CSP blocks all network except localhost, so cloud calls can only come from Rust. Sidecars: llama-helper (local summaries) and ffmpeg (import decoding). More: CLAUDE.md, docs/architecture.md.
MIT, see LICENSE.md. Upstream by Zackriya Solutions; code borrowed from whisper.cpp, Screenpipe, transcribe-rs; Parakeet by NVIDIA, ONNX conversion by istupakov.