agenius-note 0.2.0

AgeniusNote: voice-first scratchpad with local speech-to-text, optional AI parsing, quick todos, custom wake words. Cross-platform desktop app.

AgeniusNote

Talk to your computer. Get organized text. Stays on your machine.

A voice-first scratchpad for Windows, macOS, and Linux. Push-to-talk dictation, always-listening wake words, local speech-to-text (Whisper), optional AI parsing (Ollama, OpenAI, or Anthropic Claude). Quick todos in a right-side pane. A scratchpad workspace for fire-and-forget transcripts you paste elsewhere.

Local-first by default. Your voice doesn't leave the machine unless you give it a cloud key.

Previously released as voice-notes-desktop. Renamed to AgeniusNote in v0.2.0. Existing installs auto-migrate their data and keyring entries on first launch.

Why this exists

AgeniusNote started as a tool the author kept wanting and not finding. Three jobs it was built to do:

1. Capture tasks, notes, and short-term reminders without breaking flow. You're heads-down in something. A reminder hits you. Today the options are: stop, open Notion / Things / Todoist / a sticky note, type, alt-tab back. Or you swallow it and hope you remember. AgeniusNote is one click of a mic (or the wake word), 5 seconds of speech, done. The transcript is saved, the AI tags it if you let it, and you're back in the flow.

2. Be a paste-buffer between AI tools. If you work with multiple LLM chat windows open (Claude in one, ChatGPT in another, a local Ollama somewhere else), you spend a lot of time copying responses around, summarizing them out loud to yourself, then re-typing them into the next prompt. The Quick Note panel (right column, below your todos) is purpose-built for this: dictate freely, the transcript appends with each take, click Copy, paste anywhere. No DB writes, no AI parsing, no friction. Talk a thought through, hit Copy, paste the verbal sketch into your next prompt, move on.

3. Brain-dump while exploring something new. You're poking at a new app or codebase. You'd usually narrate to yourself anyway: "OK, this button does X, but when I click it I get Y, and that's weird because the docs say Z." AgeniusNote lets you actually capture that stream-of-consciousness while your hands stay on the mouse. Active Listening with a wake word means you don't even break visual focus. Later, the transcript is a step-by-step record of what you tried and what surprised you, which is gold for writing it up or filing a bug.

Built by an engineer who runs all three of these workflows daily. The design choices reflect that: keystrokes hidden behind a mic press, parsing optional and replaceable, your data on disk where you can grep it, no telemetry, MIT license.

Install

Windows

1. Download AgeniusNote-<version>-windows.zip from Releases.
2. Unzip and run agenius-note.exe. SmartScreen warns; click "More info" then "Run anyway" (we don't have a code-signing cert yet).
3. Allow microphone access when Windows prompts.
4. The first-run wizard downloads the speech-to-text model (~75 MB) so your first mic click is instant.

macOS

1. Download AgeniusNote-<version>-macos.zip from Releases.
2. Unzip, drag agenius-note.app to Applications.
3. First launch: Gatekeeper may complain; right-click then Open then confirm (we aren't notarized yet).
4. Same first-run wizard.

Linux

1. Download AgeniusNote-<version>-linux.tar.gz.
2. Extract and run ./agenius-note/agenius-note.
3. PulseAudio or PipeWire is required for mic capture.

From source (developers)

git clone https://github.com/Agenius-AI-Labs/agenius-note.git
cd agenius-note
pip install -e ".[all]"
agenius-note

Or skip the install and run the package directly: python -m agenius_note.

Requires Python 3.10+, a microphone, and ~500 MB free disk for the app plus cached models.

Features

• Push-to-talk in the Capture workspace. Click the mic, speak, body fills as soon as transcription returns. AI parses out title / tags / priority / type in a second pass without blocking the UI.
• Active Listening with a wake word (default: hey_jarvis). Say the phrase, dictate, app auto-stops on silence.
• Quick Note panel in the right column (under your todos). Fire-and-forget transcripts. No DB writes, no parsing. Talk, copy, clear. When the panel has focus, Active Listening transcripts land there raw instead of in Capture.
• Quick Todo pane on the right. Type or say "quick todo: water the plants" and it lands in the running todo list. Check off items as you go.
• Local speech-to-text via faster-whisper. Auto-detects CUDA for GPU acceleration (10-30x realtime on a recent Nvidia card). Falls back to CPU.
• Optional AI parsing via local Ollama, OpenAI's gpt-4o-mini, or Anthropic Claude Haiku. Or none, if you just want raw transcripts.
• Custom wake words via openWakeWord. Train your own phrase in Colab, drop the .onnx into Settings.
• Themed UI in PySide6: dark, light, and cyberpunk. Inter + JetBrains Mono.

How it works

Three workspaces plus a persistent right-side column:

Surface	What it does
Capture	Voice or typed input, AI-parsed into structured notes/tasks, saved to SQLite.
Tasks	List of saved tasks. Click to edit.
Notes	List of saved notes. Click to edit.
Right column (top half)	Quick Todos pane. Add via text or voice ("quick todo: ..."). Check off when done.
Right column (bottom half)	Quick Note panel. Voice or typed scratchpad. Transcript appends with each take. Copy or Clear. Nothing persists.

Custom wake words: safety note

The "Custom model file" setting in Settings then Active Listening loads a .onnx or .tflite file through onnxruntime. Only load files you trained yourself (via the openWakeWord Colab notebook) or downloaded from openWakeWord's official model zoo. A malicious .onnx file from an untrusted source could exploit a future onnxruntime vulnerability.

Privacy

• Speech-to-text runs locally. No audio leaves your machine.
• AI parsing is off by default. Enable Ollama for local LLM, or OpenAI / Anthropic for cloud. Skip all three for raw transcripts.
• No telemetry. The app makes one network call on first launch: downloading the Whisper model from Hugging Face Hub. After that, fully offline-capable.
• Your data lives in your OS's standard user-data location (see Deployment). Delete the folder to delete everything.

Documentation

• Architecture, signal flow, threading, schema.
• Deployment, install paths in depth, where data lives, GPU setup.
• Custom wake words, train your own phrase via Colab.
• Contributing, dev setup, code style, PR process.

Stack

PySide6 · faster-whisper · openWakeWord · Pillow · OpenAI / Anthropic / Ollama (optional) · SQLite

Status

Beta. Daily-driven by the author. APIs may shift between v0.x releases. v1.0 lands once the install paths are stable across all three OSes and the test coverage is meaningful.

License

MIT. See LICENSE.

Acknowledgments

Built on top of excellent open-source work. Particular thanks to:

• faster-whisper for CTranslate2-backed Whisper inference.
• openWakeWord for free, trainable wake-word detection.
• PySide6 for the cross-platform UI.

Created by Michael Frostbutter under Agenius AI Labs.