Push-to-talk voice
Ctrl+K starts a faster-whisper voice loop with interruptible LLM streaming and sentence-level Piper TTS.
A frameless animated character that listens for your voice, answers with neural TTS, remembers what matters, and can run cloud-fast with Groq or locally through Ollama and Piper fallbacks.
“You mentioned preferring espresso. Want me to remember that?”

KIBO is a desktop-resident companion: visible as a character, reachable through hotkeys, backed by a local memory vault, and governed by policy when it speaks first.
Ctrl+K starts a faster-whisper voice loop with interruptible LLM streaming and sentence-level Piper TTS.
Facts are embedded with sqlite-vec and also written to an Obsidian-compatible vault you can inspect and edit.
A frameless transparent companion uses VP9 alpha WebM states for idle, thinking, talking, acting, and happy moments.
Morning greetings, idle check-ins, battery alerts, and meeting reminders obey quiet hours, snooze, and daily caps.
Each dependency has a graceful degradation path so the app can still start when keys, models, or network access are missing.

First-run wizard explains provider choice, audio, privacy, hotkeys, and tray controls.
Memory and proactive behavior are explicit opt-ins with settings to inspect, edit, delete, clear, snooze, or disable.
Diagnostics export redacts sensitive data by default and keeps logs under the local ~/.kibo directory.
Install with uv, optionally add Groq and Piper, then launch the Python backend and Electron frontend.
git clone https://github.com/yash-dev007/KIBO.git
cd KIBO
pip install uv
uv sync
# Windows launcher
KIBO.batBackend API runs on port 8765.
Local data lives under ~/.kibo.