Private
Stays on your LAN
Audio, transcripts, and model calls can all run inside your own network. No third party in the loop unless you put one there.
// Self-hosted · Open source · v3.0.0
Home voice AI that runs on your hardware.
Kenzy is a distributed, self-hosted voice assistant built from six independent microservices — wake word, speech-to-text, a language model, speaker ID, and text-to-speech. Point the LLM at a model on your own machine and your voice never leaves the network.
$ curl -fsSL https://kenzy.dev/install.sh | bash
// one-line installer landing soon — manual setup on Get Started
6
Microservices
100%
Local-capable
~$0
Per-request cost
Pi Zero 2 W
Runs on edge
01 Why local
Your voice never has to leave the house.
Most assistants ship every word you say to a server you don't control. Kenzy flips that: because kenzy-llm runs on LiteLLM, you can point it at a model running on your own box — Ollama, LM Studio, vLLM — or a cloud provider if you'd rather. Your call, per service.
Open model
Bring your own LLM
LiteLLM speaks to local runtimes and every major provider. Swap models with one line of YAML — no rewiring.
No meter
No per-word bills
Run it on hardware you already own and the marginal cost of "what's the weather?" is electricity, not API credits.
Hackable
Yours to take apart
Plain Python, readable configs, and a one-file skill system. Built to be tinkered with, not locked down.
02 What's inside
A real assistant, assembled from parts you can see.
kenzy-node
On-device wake word
openWakeWord runs on every frame locally, with an optional Silero VAD gate to kill false triggers. Train and drop in your own wake word.
kenzy-speaker
Knows who's talking
SpeechBrain ECAPA-TDNN identifies enrolled speakers — so unlocking the front door by voice can require a recognized person.
kenzy-llm
Skills & tool-calling
Drop an async function in skills/, decorate it with @skill, and the model calls it as a tool. No registration, no boilerplate.
fast path
Instant commands
Common phrases like "turn on the lights" resolve deterministically — no model round-trip — so they answer the moment you finish speaking.
v3.0.0 · GROUND-UP REWRITE
Rebuilt from the ground up.
Kenzy v3 is a complete redesign — not a refactor. The monolith is gone, replaced by six small services that each do one job and talk over a simple WebSocket + PCM protocol.
The result is a system you can spread across the house: a featherweight node on a Raspberry Pi Zero 2 W in each room, the heavy lifting on a server or workstation wherever you've got the horsepower.
- More flexible — run every service on one box or scatter them across the network.
- More responsive — a deterministic fast path skips the LLM for everyday commands.
- Edge-ready — nodes do only wake word + capture, so they fit tiny hardware.
# point Kenzy at a model on your own machine model: "ollama/llama3.1" base_url: "http://localhost:11434" # ...or a cloud provider, same two lines # model: "gpt-4o" # model: "claude-opus-4-8"
Ollama
LM Studio
vLLM
OpenAI
Anthropic
any LiteLLM provider
03 The stack
Six services. One pipeline.
ON DEVICEnodeWake word, capture & playbackkenzy-node
PORT 8765serverWebSocket hub & orchestratorkenzy-server
PORT 8767sttSpeech-to-text · faster-whisperkenzy-stt
PORT 8766llmLLM & skills · LiteLLMkenzy-llm
PORT 8769ttsText-to-speech · OpenAI / Kokorokenzy-tts
PORT 8768speakerSpeaker ID · SpeechBrainkenzy-speaker
// Bring it home
Stand up your own voice assistant.
Clone it, install the services you need, point it at your hardware, and start talking. The docs walk you through every step.