- Category
- AI Tools
- Rank
- No. 744Tools index
- Listed in
- #2 Capture meetings
- Pricing
- Open Source
- Platform
- desktop
- Type
- APP
- Builder
- @so_ainsight
- GitHub
- 30.0k stars
- Latest release
- v0.4.0
- Date
About
Meetily is a privacy-first AI meeting assistant that transcribes, diarizes, and summarizes meetings entirely on your local machine using Whisper/Parakeet and Ollama — no cloud, no data leakage. It runs on macOS and Windows with GPU acceleration and supports flexible AI providers including local Ollama models.
What it does
Meetily records a meeting from your microphone and the machine's own output at the same time, turns the audio into text on the device using a Whisper or Parakeet model, then writes structured notes with a language model you choose. It ships as one desktop program: a Rust core handles capture, speech to text, storage in a local SQLite file and summary orchestration, while a Next.js interface renders the transcript and an editable summary. The summary step can point at a locally hosted model or a hosted provider you supply a key for.
Why it's ranked here
The privacy claim survives contact with the source. Audio capture, speech to text and storage all sit in the Rust core, and usage analytics are opt in and off by default. Two things temper the verdict. The documentation doubles as a sales page for a paid edition built on a separate codebase, complete with a discount code, and speaker separation, the feature meeting notes most need, appears only as a planned item for that paid tier. A Python service also lingers in the tree, flagged by its own notes as archived, unsupported, and not to be treated as a production API.
What's good
The audio side is where the effort went. Capture covers microphone and system output together, with voice activity detection, a mixer, device disconnect monitoring, buffer pooling and checkpointed saving during long recordings. Model choice is genuinely open: a local host, Anthropic, Groq, OpenRouter, OpenAI, or any OpenAI-compatible endpoint where you set base URL, key, model, token cap, temperature and top-p sampling. Builds select an accelerator per platform, covering Metal and CoreML on Apple silicon and CUDA, Vulkan, HIPBlas or OpenBLAS elsewhere. Windows and macOS get prebuilt installers.
Tradeoffs
Linux users build from source, and acceleration needs the development toolkit installed, not just a GPU driver: the build guide is explicit that drivers alone fall back to CPU, and Vulkan additionally needs two environment variables set. The free edition has no speaker labels, no automatic meeting detection and no formatted export, all of which sit behind the paid tier. The archived Python service still in the tree once exposed an API that allowed every origin with no authentication, so anyone skimming the repository could wire up something they should not. The privacy policy carries an unfilled date placeholder.
How to use it well
Best fit is one person on a laptop who cannot let client, patient or legal conversations reach a vendor server, and who will run a local model host for the summary step. It earns a slot as the recorder and note taker beside any call, since it listens to system output rather than joining as a bot participant. It does not cover team workflows: no shared meeting library, no calendar hookup, no automatic joining, and no speaker attribution in the open edition. Pair it with your own storage and sharing, and expect a build step on Linux.
Technical notes+
The Tauri entry module frontend/src-tauri/src/lib.rs declares the module graph (audio, transcription engines, summary, database, analytics, notifications, plus one provider module each for local hosting, OpenAI, Anthropic, Groq and OpenRouter) and exposes recording start and stop, device listing, level monitoring, language preference and transcript saving as Tauri commands guarded by an atomic recording flag. frontend/src-tauri/src/audio/mod.rs is the widest surface: it re-exports capture, pipeline, stream, an FFmpeg-style mixer, voice activity detection, level monitoring, buffer pooling, incremental saving, device monitoring, retranscription and import submodules. frontend/src-tauri/src/summary/mod.rs defines CustomOpenAIConfig with endpoint, optional API key, model, max tokens, temperature and top-p, and re-exports chunking helpers plus template commands. frontend/src-tauri/src/whisper_engine/mod.rs splits engine, acceleration, system monitoring and a parallel processor. frontend/src-tauri/src/database/mod.rs and frontend/src-tauri/src/ollama/mod.rs are thin module roots over repositories and provider code. frontend/package.json pins Next.js and React with one Tauri script per acceleration feature; docs/architecture.md and docs/BUILDING.md describe the detection order that picks among them. backend/app/main.py still sets CORS to a wildcard origin with credentials allowed, and backend/README.md marks that whole directory archived and unsupported. LICENSE.md is MIT and PRIVACY_POLICY.md documents opt-in PostHog analytics.
Observed
- License
- MIT, copyright Zackriya Solutions (LICENSE.md)
- Stack
- Rust core inside a Tauri shell with a Next.js and TypeScript interface
- Interface
- Desktop application only; no CLI or importable library in the supported path
- Platforms
- Prebuilt installers for Windows and macOS; Linux is built from source
- Transcription
- Local Whisper or Parakeet models, no cloud call required
- Storage
- Local SQLite database for meeting metadata, transcripts and summaries
- Summary providers
- Local host, OpenAI, Anthropic, Groq, OpenRouter, or any OpenAI-compatible endpoint
- Acceleration
- Build features for Metal, CoreML, CUDA, Vulkan, HIPBlas and OpenBLAS
- Telemetry
- PostHog usage analytics, opt in and off by default per PRIVACY_POLICY.md
- Legacy code
- backend/README.md marks the Python FastAPI service archived and unsupported
- Commercial tier
- README.md advertises a paid edition on a separate codebase for accuracy, exports and diarization
Read from README.md, LICENSE.md, PRIVACY_POLICY.md, docs/architecture.md, docs/BUILDING.md, frontend/package.json, backend/requirements.txt, backend/app/main.py, backend/README.md, frontend/src-tauri/src/lib.rs, frontend/src-tauri/src/audio/mod.rs, frontend/src-tauri/src/whisper_engine/mod.rs, frontend/src-tauri/src/summary/mod.rs, frontend/src-tauri/src/database/mod.rs, frontend/src-tauri/src/ollama/mod.rs.
What it can do
Transcribe meeting audio in real-time
Live microphone or system audio capture → Real-time text transcript of spoken words
Identify and label individual speakers in a meeting
Audio stream with multiple participants → Diarized transcript with speaker-attributed segments
Summarize meeting content using a local AI model
Meeting transcript text → Condensed meeting summary with key points
Run speech-to-text transcription entirely on-device
Audio input and local Whisper or Parakeet model → Transcribed text with no data sent to external servers
Accelerate transcription and AI inference using GPU
Audio or text input with available GPU hardware → Faster transcription and summary generation via GPU acceleration
Generate meeting summaries using a locally hosted Ollama model
Meeting transcript and a configured Ollama model → AI-generated summary produced without cloud API calls
Capture system or microphone audio during a meeting
Active meeting session on macOS or Windows → Raw audio stream ready for transcription processing
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.