Skip to content
Arthur-101Public

About

No description, website, or topics provided.

Resources

Stars

2 stars

Watchers

0 watching

Forks

Repository files navigation

Aether — local-first AI for Windows: pill, peek, island.

Ask anything. Drop a file. Point at the screen.
Windows 10/11 · x64 · Tauri + Python · Apache-2.0 · no account required

Latest release Apache-2.0

Aether combines a voice-driven Windows desktop assistant with a zero-friction Drag + Shift radial file converter and a top-anchored "Desktop Island" overlay.

Aether in action: the Desktop Island, Quick Chat, and the Drag + Shift radial file wheel.

Aether_video.mp4

▶ Watch the 40-second tour — the Island, Quick Chat and the Drag + Shift file wheel, rendered from the app's own React components rather than a mock-up.
Music: “Happy Beats Business Moves Vol. 11” by Sascha Ende (ende.app), CC BY 4.0.


A closer look

The Desktop Island: controls, audio, display, connectivity, widgets and a colour picker in one panel.
The Desktop Island — controls, audio, display, connectivity, widgets and a colour picker, one hover away.

Quick Chat Drag + Shift
Quick Chat reading a document and answering in place. The radial wheel offering crop, compress, convert and Ask AI for a PDF.
Ask about the file in front of you: it reads it, runs the tool, and answers in place. Drop a file on Explorer and hold Shift to pick a conversion or edit.

Architecture Highlights

  • Frontend: Tauri v2 (Rust backend + React 18 + TypeScript + Tailwind CSS + Framer Motion) always-on-top transparent overlay (< 40MB RAM).
  • Backend: Python 3.12 async daemon (FastAPI + WebSockets + FastMCP).
  • Local LLM: Optional engine packs drive the bundled llama-server.exe from the PrismML llama.cpp fork — Vulkan by default, and never the CPU. The suggested Ternary-Bonsai-2-9B-PQ2_0-MTP-Q8_0.gguf needs that fork's ternary PQ2_0 (ggml type 142) support. Nothing is bundled: install an engine pack, then point Settings at a .gguf.
  • Memory & Context Pipeline: Provider-aware memory depth (local default 2 recent turns + 3 semantic matches; cloud 8 + 8) plus session-scoped vector search via sqlite-vec & CPU fastembed (bge-small-en-v1.5, 384 dims, 0 VRAM).
  • Voice Stack: Local Phonon-2 STT + 24kHz ONNX TTS (Kokoro/Paradee families) with voices discovered from the model's sidecar.
  • Deterministic Tooling: Typed FastMCP tools for image/document conversions, metadata stripping (piexif), and APScheduler reverse-trigger loops.
  • Zero Hardcoded Paths: Dynamically resolves %APPDATA%/Aether and user-configurable paths via config_manager.py.

Directory Structure

aether/
├── bin/                          # Portable binaries (ffmpeg.exe; engine packs add llama-server.exe)
├── backend/
│   ├── config_manager.py         # Portable config manager (%APPDATA%/Aether/config.json)
│   ├── requirements.txt          # Python dependencies
│   ├── test_phase1.py            # Phase 1 verification test
│   ├── db/
│   │   ├── database.py           # SQLite + sqlite-vec setup & tables
│   │   └── sessions.py           # Provider-aware recent window & session-scoped vector memory
│   ├── ai/
│   │   ├── embedder.py           # CPU fastembed (bge-small-en-v1.5)
│   │   ├── llm_router.py         # Local llama-server / Ollama / cloud router
│   │   ├── stt_engine.py         # Phonon-2 STT engine
│   │   ├── tts_engine.py         # Family-detected local ONNX TTS engine
│   │   └── local_voices.py       # Voice-sidecar discovery + phoneme tokenization
│   ├── mcp_tools/                # FastMCP deterministic tools & APScheduler
│   └── os_hooks/                 # Windows OS hooks (pynput, win32com, pywinauto, winsdk)
└── frontend/                     # Tauri v2 + React overlays (DesktopIsland, RadialWheel, Drawer)

Setup & Verification

# 1. Install backend dependencies (Python 3.12 recommended)
cd backend
pip install -r requirements.txt

# 2. Install the isolated Phonon-2 speech-to-text runtime
cd ..
powershell -ExecutionPolicy Bypass -File setup_stt.ps1

# 3. Run the full backend suite
py -3.12 -m pytest backend -q

# 4. Build the frozen daemon (PyInstaller onedir, Python 3.12)
powershell -ExecutionPolicy Bypass -File build_sidecar.ps1

# 5. Build the production overlay and installers
cd frontend
npm run tauri build

The build writes Aether_<version>_x64-setup.exe (currently 1.0.3) under frontend/src-tauri/target/release/bundle/nsis/. v1.0 ships NSIS only; the WiX MSI was built and verified during 0.9 but is not part of the release target. The installer carries the overlay plus the frozen daemon at aether-backend\; the isolated STT runtime remains an optional stt-phonon2 pack (see PACKAGING_AND_CONFIG.md). Building the sidecar needs PyInstaller >= 6.22 — 6.12 cannot freeze numpy 2.5.

Phonon-2 runs in an isolated Python runtime — the optional stt-phonon2 pack, or a dev checkout's bin/stt-runtime — so Torch and Transformers never load in the backend process. Without it, dictation stays disabled and logs one actionable warning; the rest of Aether is unaffected. Install it from Settings → Components, or point Settings → Voice → STT Model at a directory containing model.fermion and config.json.

Existing configs are upgraded in place with any newly added defaults. An unreadable config.json is preserved as config.corrupt-<timestamp>.json before Aether restores safe defaults.

Aether keeps rotating diagnostics under %LOCALAPPDATA%\Aether\Logs: aether-backend.log (written by backend/logging_setup.py) and aether-overlay.log (written by the Tauri overlay). Launching Aether twice focuses the running overlay instead of starting a second one.


Licence

Aether is free and open source under the Apache License, Version 2.0. You may use, modify, and redistribute it, including commercially, provided you keep the copyright and licence notices.

Third-party components keep their own licences — pynput is LGPL-3.0, certifi and tqdm are MPL-2.0 — and every one of them is listed with its licence text in THIRD_PARTY_LICENSES.txt. See NOTICE for the attribution summary.

Optional downloads are separate works with their own terms: the espeak-ng voice pack is GPL-3.0-or-later, the ffmpeg media pack is LGPL-3.0-or-later, and model files (LLM, STT, TTS voices) carry the licence of the model you choose.

About

No description, website, or topics provided.

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages