Meetily

Meeting notes without a bot joining the call

Meetily is a desktop app that records a meeting, transcribes it as it happens and then writes up the notes with a language model. No bot joins the call. Meetily records the microphone and the system audio, so it does not care which meeting software is on the other end: Zoom, Google Meet, Teams or anything else. It is written in Rust on Tauri and keeps meetings and transcripts in a local SQLite database.

Murmure runs the same Parakeet v3 model, but for dictation, for example into an AI assistant. Scriberr transcribes audio files from a self-hosted server. The comparison of local speech-to-text tools sets out what each one asks of your hardware and what leaves the machine.

Pick Whisper or Parakeet, summarize with built-in llama.cpp

• Two families of transcription models: Whisper through whisper.cpp, from tiny up to large-v3 (large-v3-turbo is the default), or NVIDIA's Parakeet TDT 0.6B as an int8-quantized ONNX model.
• Local summaries without installing Ollama: the app bundles llama.cpp and lists four GGUF models, Qwen 3.5 at 2B and 4B, Gemma 3 at 1B and 4B. Gemma 3 1B is listed at about 1 GB of RAM, Gemma 3 4B at about 3.5 GB. Ollama and any OpenAI-compatible endpoint still work.
• Re-transcribing a recording (beta): an imported audio file or a past meeting can go through a different model or language, still on your machine.
• Notes in a different language from the meeting: since v0.4.0, the summary can come out in any of 30 listed languages, whatever language was spoken.
• GPU acceleration out of the box: Metal and Core ML in the macOS package, a Vulkan build of Whisper in the Windows installer.

Speaker identification costs extra, and Linux needs a source build

Audio stays on the machine; text can leave it. Transcription runs locally. For the summary, the text goes over the network as soon as you pick a remote provider: Claude, Groq, OpenRouter, OpenAI or any OpenAI-compatible endpoint. Meetily's pricing page states that the provider gets the transcript text, never the audio. The app also goes online to fetch models, from Hugging Face for Whisper and the summary models, and from meetily.towardsgeneralintelligence.com for Parakeet v3. PostHog analytics have been off by default since v0.4.0; when enabled, they send usage metrics (models, durations, audio device type) and no meeting content.

No Linux or Intel Mac package. Each release ships a Windows x64 installer and an Apple Silicon (aarch64) macOS image, nothing else. On Linux you build an AppImage from source with Rust, Node.js and pnpm. The Windows installer needs an x64 CPU with AVX2, and CUDA is only available if you compile it yourself. On an Intel Mac, BetterWispr, a dictation app that also takes meeting notes, ships its own Intel build.

The free edition does not tell speakers apart. The GitHub repository description mentions speaker diarization, yet the README lists speaker identification as a PRO feature. In the Community Edition code, each transcript segment only records its source, mic or system. Scriberr does separate the voices in a recording.

PRO costs $10 per user per month. It is billed yearly, sold under a commercial license and built on a separate codebase. It adds speaker identification, custom summary templates, PDF, DOCX and Markdown exports, automatic meeting detection, and self-hosted deployment for a team. The README aims it in particular at teams of 2 to 100 people and points larger organizations to an Enterprise tier with custom pricing. The same README says the MIT-licensed Community Edition will stay "free & open source forever".

Still at version 0.x. Four releases came out in 2026; the latest, v0.4.1 on September 12, is mostly bug fixes. On October 7, 2026, the repository had 202 open issues and 158 open pull requests.