Platypus Notes

An open-source desktop app for taking notes, transcribing meetings, and chatting with your documents. Runs entirely on your machine — the only network calls are to whichever LLM you choose to connect, and only when you ask a question.

Download for macOS · platypusnotes.com · MIT license

Platypus_Github.mp4

What it does

• Capture meetings — auto-detects Zoom and Teams calls; transcribes locally via Whisper or via OpenAI's API
• Organize notes and documents — rich editor with PDF/DOCX/TXT import, project grouping, full-text search, and AI-assisted polish
• Chat with everything you've written — per-project HNSW vector search, with Claude, OpenAI, Gemini, or any local Ollama model. Answers are grounded in your notes with inline [n] citations you can click to jump to the source passage, and each project greets you with suggested questions generated from its content
• Generate from any note — turn a meeting transcript or document into structured meeting notes with decisions and action items, a follow-up email, a slide deck, or an audio podcast (via ElevenLabs)

Data stays on disk in SQLite. In local transcription mode, audio never leaves your machine.

How it compares

	Platypus	Granola	NotebookLM	Otter.ai
Stores data on your machine	✅	❌	❌	❌
Meeting transcription	✅	✅	❌	✅
RAG over your notes/docs	✅	partial	✅	❌
Bring your own LLM	✅	❌	❌	❌
Open source	✅	❌	❌	❌
Free	✅	partial	partial	❌
Native desktop	✅	✅	❌	❌

Voice transcription

Two modes, switchable in Settings.

Local Whisper (default) — on-device transcription via whisper.cpp.

• Real-time: live transcript streams during recording
• Works offline, no API key required
• Hardware-accelerated via Metal on macOS, CPU fallback elsewhere
• Models (selectable in Settings): Large v3 (~3.1GB, default, best quality), Large v3 Turbo (~1.6GB), Distil Large v3.5 (~1.5GB, fastest)
• Model auto-downloads on first use

OpenAI API — records WAV, uploads to OpenAI's Whisper endpoint.

• Requires an OpenAI API key
• Transcribes after recording finishes (not real-time)

Tech stack

Layer	Technology
Desktop shell	Tauri v1 (1.5.2)
Backend	Rust
Frontend	React + TypeScript + Vite
UI	Chakra UI + styled-components
Editor	TipTap
AI providers	Claude, OpenAI, Gemini, Ollama
Transcription	whisper-rs v0.16 (local) / OpenAI Whisper API (cloud)
Audio	CPAL (recording), nnnoiseless (denoising), rubato (resampling)
Database	SQLite (rusqlite)
Vector search	HNSW (hnswlib-rs)

Build from source

Requirements

• Node 18+ (recommended via nvm)
• Rust
• cmake — required by whisper-rs-sys to compile whisper.cpp
  • macOS: brew install cmake
  • Windows: winget install Kitware.CMake
• LLVM / libclang (Windows only — required by bindgen when building whisper-rs-sys; macOS ships this via Xcode Command Line Tools)
  • winget install LLVM.LLVM
  • Then set LIBCLANG_PATH so bindgen can find libclang.dll:

    setx LIBCLANG_PATH "C:\Program Files\LLVM\bin"
    

    Open a new terminal afterward so the env var is picked up.

Run in dev

npm install
npm run tauri dev

If you hit dependency issues, delete package-lock.json and re-run npm install.

Add your LLM API keys in Settings before use.

Build a release

npm install
npm run tauri build

For a signed + notarized macOS build that uploads to your S3 bucket, see scripts/build-mac.sh — requires Apple Developer credentials in .env.build.

Architecture notes

A few of the less-obvious decisions:

• Audio pipeline: CPAL capture → energy-based VAD on raw samples → nnnoiseless denoising at 48kHz → rubato resample to 16kHz → whisper.cpp via whisper-rs. VAD runs before denoise because RNNoise crushes signal amplitude ~100x and every chunk would otherwise look silent.
• Meeting detection: Zoom is detected by presence of the CptHost process; Teams by CPU usage on its audio.mojom.AudioService sub-process. No Zoom/Teams API access required.
• Vector search: per-project HNSW indices (hnswlib-rs); documents chunked and embedded on save when vectorization is enabled.
• Grounded chat with inline citations: retrieved chunks are numbered [1]..[n] in the system prompt, the LLM is instructed to cite every claim and refuse to invent answers, and the same numbering is emitted to the UI so each [n] in the response is a clickable chip that opens the supporting passage.

Contributing

See CONTRIBUTING.md.

Acknowledgments

Platypus stands on the shoulders of:

• whisper.cpp and whisper-rs — local speech-to-text
• Distil-Whisper by HuggingFace — the distilled Whisper variants
• nnnoiseless — pure-Rust port of Mozilla's RNNoise
• rubato — sample-rate conversion
• hnswlib-rs — HNSW vector index

License

MIT.