Duplex voice engine
A "Hey Orion" keyword spotter on a circular mic buffer, Silero VAD, Faster-Whisper, and Kokoro-82M sentence streaming. Talk over it and playback flushes in one 20 ms block.
Python 3.11+ · macOS · Windows & Linux beta
Build Your Own Local Voice AI. Talks Back in ~2 Seconds. Offline. Powered by Nebula Agents.
A production-grade Python starter kit with low-latency duplex audio, multi-model routing (Ollama, Gemini, Claude), and MCP desktop automation.
- ✓ Runs fully offline on Ollama
- ✓ "Hey Orion" wake word + barge-in
- ✓ Instant license key
orionkit — python main.py
$ python main.py
● nebula MCP up · 4 tools · router: ollama ✓ gemini ✓ claude ✓
👂 wake hey orion
you turn off the kitchen lights and tell me my RAM usage
· intent vision · tool_action → gemini-flash
· tool home_assistant(toggle light.kitchen) ✓
· tool system_telemetry() ✓
orion
Latency breakdown
Speak. Hear an answer. No awkward pause.
elapsed
0 ms
OrionKit pipeline
speaking
Same models, no streaming
6.4 s
transcribe → wait for the whole reply → synthesize all of it → play
Measured medians on an Apple M1 Pro (32 GB) under heavy load, fully local with qwen2.5:3b, using tools/bench_latency.py, from the end of your sentence to the first spoken word, long answer ("Explain how a rainbow forms"). Small talk starts in about 1 s. The "Hey Orion" wake word fires about 0.2 s after you say it. Faster hardware and cloud models change these numbers.
What's inside
Three layers you'd otherwise spend months on
Tri-model router
Local intent classification picks fast (Ollama), vision (Gemini) or deep (Claude). Circuit breakers route around 429s, 5xx, empty credit and dropped Wi-Fi without throwing.
Nebula MCP agents
A Model Context Protocol dispatcher with a sandboxed shell, allowlisted files, host telemetry and Home Assistant, plus any MCP server you add.
Free core vs Pro
Start free. Ship with Pro.
| Feature | Open Source Core Free · MIT | OrionKit Pro $99 one-time |
|---|---|---|
| Clone core |
One price. Yours forever.
Get OrionKit Pro
Pay once, get the full source and a license key the moment checkout completes. The key is verified offline, so your assistant never phones home.
- 1Checkout in a secure overlay. Card, Apple Pay, Google Pay.
- 2Your signed license key appears right here. Save it; you can recover it anytime with your email and order number.
- 3Run
python main.py --license <key>once. Done.
OrionKit Pro
$99
one-time · lifetime updates · taxes calculated at checkout
Payments handled by our merchant of record. 7-day refunds. License.
Already bought OrionKit Pro?
Recover your license
Enter the email you paid with and the order number from your Lemon Squeezy receipt (for example #123456).
FAQ