A self-hosted home assistant you talk to

How a spoken command is answered

Each step runs only when the one before it could not answer. A light switched by voice never waits on a model, a plugin or a cloud call it does not need.

1. 01

   Heard in the room

   The satellite hears "Niles" itself, with a wake word model trained for that one word. Nothing streams until it is said, and the room's Sonos turns down while Niles listens.

   The satellite

2. 02

   Fixed phrasings first

   Lights, blinds, scenes, music, the TV, timers, heating, the weather and the clock are matched by regex and carried out with no language model. The target is about half a second end to end.

   Tier 0

3. 03

   A fast model for the rest

   Anything else goes to a fast model (Cerebras by default) with only the tools that request is likely to need. You can say it your own way, and it takes a second or so.

   Tier 1

4. 04

   The assistant for bigger jobs

   Making a rule, writing a travel plan or a podcast, or reading a photo goes to Claude Code or ChatGPT on the household's own subscription. It reaches the same tools through Niles's own MCP server.

   The assistant

5. 05

   Answered where you asked

   Piper speaks the reply on the satellite that heard you. A reply that ends in a question keeps the microphone open, so you can answer without the wake word.

   The satellite

Talking to Niles lists what you can say, and the voice loop follows a command through each tier and the prompt the model gets.

Why Niles says how it is built, what leaves the house, and whether it fits yours.

In the app

What you say to a satellite is in your chat too, and you can type there instead. Ask for a plan, a trip or a document and the assistant writes it in the background and keeps it under Saved, with every version.

Run it yourself

Niles is one container and a Postgres database, beside the MQTT broker and Zigbee2MQTT you run for your Zigbee devices. It reaches the database through NILES_DATABASE_URL and seals the keys you type into the app with NILES_SECRET_KEY. The first start also needs the broker's login and the ElevenLabs and Cerebras keys. Everything else is set in the app.

git clone https://github.com/MarkNygaard/niles
cd niles/deploy/compose
cp .env.example .env
# In .env: the database password, the secrets key,
#   the broker's login and the two API keys
# In niles.toml: your MQTT broker's address

docker compose up -d

git clone https://github.com/MarkNygaard/niles
cd niles
cp -r deploy/kustomize/overlays/dev deploy/kustomize/overlays/home
# In overlays/home: the broker's address in niles.toml,
#   and the namespace and image tag in kustomization.yaml

kubectl create namespace niles
kubectl create secret generic niles-secrets \
  --namespace niles \
  --from-literal=database-url='postgres://niles:...@postgres:5432/niles' \
  --from-literal=secret-key="$(openssl rand -base64 32)" \
  --from-literal=mqtt-username='...' \
  --from-literal=mqtt-password='...' \
  --from-literal=elevenlabs-api-key='...' \
  --from-literal=cerebras-api-key='...'

kubectl apply -k deploy/kustomize/overlays/home

The image is ghcr.io/marknygaard/niles, for x86-64 Linux. Pin its tag to choose when you upgrade. Self-hosting covers the first sign-in, backups, the ports and HTTPS for phones.

Built in

Each of these is a crate in the repository, held to its tests. Anything else is an add-on.

Zigbee2MQTT

lights, plugs, sensors, switches and blinds, named room/device

WLED

LED strips, as a second MQTT source

Sonos

music and radio in each room, and Spotify through it

LG webOS

TVs over their local websocket, woken by Wake-on-LAN

tado° and UniFi

the heating, and who is home

Microsoft, Google, iCloud

each person's calendars and mail

Strava

each person's workouts, from their own Strava app

Linear

issues and projects, through Linear's own MCP server

SearXNG

the house's own web search, for the assistant