Kåre The Smart Butler

The AI layer your smart home is missing.

Frigate watches your cameras. Home Assistant controls your devices. KTSB is the intelligence that ties them together — running entirely on your own hardware, with no cloud required.

---

Why does this exist?

If you run a smart home with Home Assistant and Frigate, you have probably noticed something: adding AI to the mix destroys the machine it runs on.

It makes sense when you think about it. Home Assistant is an I/O powerhouse — juggling hundreds of sensors, automations, and integrations all day. Frigate is doing real-time video processing for every camera on your property. Neither of them has spare capacity for a large language model. No matter how powerful the hardware you buy, cramming all three onto one machine is a death sentence.

KTSB is the third piece. It runs on its own dedicated machine — even a mid-range mini PC works — connects to your existing Home Assistant and Frigate installs, and handles everything AI-related:

• Understanding what you say in natural language
• Controlling your smart home through Home Assistant
• Analysing camera footage through Frigate
• Answering questions, setting timers, searching the web
• Remembering your preferences and conversations over time

Each system stays focused on what it does best.

---

What can Kåre do?

Talk and understand

Ask Kåre anything in natural language — by voice or text. No rigid commands, no keyword memorisation. "Turn off everything downstairs and lock the door" works just as well as "lock front door".

The GUI supports Norwegian, English, and German. The assistant name, hot-word, and response language are all configurable.

Control your smart home

Kåre connects to Home Assistant and can control anything HA knows about: lights, locks, thermostats, covers, media players, scenes, and scripts. You define your own room names and device nicknames so you can say "the lamp in the corner of the bedroom" and Kåre knows exactly what you mean.

See through your cameras

Ask "is there anyone outside?" or "what happened at the front door in the last hour?" and Kåre checks your Frigate cameras and describes what it sees. Motion events and snapshots can be analysed automatically and announced over your speakers.

Remember and learn

Kåre maintains a short-term memory of your recent conversations and a long-term memory that persists across sessions. It builds a self-image over time based on its interactions — the longer you use it, the better it knows you.

A team of AI assistants

Behind the scenes, Kåre has a small team working alongside it:

• Miss Kåre — a quiet evaluator who reviews Kåre's answers before they reach you
• Mechanic — a technical assistant for deep-dives, system inspection, and complex problems
• Miss Library — searches a local Wikipedia corpus and the web for factual answers
• Jing and Jang — background inner voices that process and reflect while you are not watching

Every night at 04:00 they hold a reflection meeting to review the day. At 05:30 a developer meeting investigates any system issues. All of this happens quietly while you sleep.

Multi-user with proper access control

Create separate accounts for everyone in your household. Each user gets a PIN, a role (Child, Teen, Adult, Admin), and their own memory and conversation history. Sensitive tools — web search, system inspection, camera access — are gated by role.

Remote access via WireGuard VPN is available as an optional service, with per-user access levels: local only, AI chat only, or full smart home access from anywhere.

Flexible AI backend

Kåre uses Ollama for local LLM inference out of the box. No cloud account required. If you have a capable GPU, models run fast and privately on your own hardware. If you do not have a GPU, or want a more powerful model for complex reasoning, you can connect any OpenAI-compatible provider: OpenAI, NVIDIA NIM, vLLM, HuggingFace, or others.

---

What you need

Kåre runs via Docker, which means it works on Linux, Windows, and macOS without installing anything else manually.

	Minimum	Recommended
OS	Linux, Windows 10/11, macOS	Linux (native Docker, fastest)
RAM	8 GB	16 GB+
Disk	20 GB free	60 GB+ (for AI models)
GPU	Not required	NVIDIA with 8 GB+ VRAM

No GPU? Kåre still works fully. Connect a cloud LLM provider (like OpenAI) in Settings, or run small models on CPU — slower but functional.

No Home Assistant or Frigate? That is fine too. Kåre works as a standalone AI assistant and you can add integrations whenever you are ready.

---

Getting started

Step 1 — Install Docker

Docker is a tool that packages software into self-contained boxes that run the same way on any computer. You do not need to install Python, databases, or any other programming tools — Docker handles all of it.

Windows: Download and install Docker Desktop. During installation it will ask you to enable WSL 2 — say yes. Restart your computer when prompted. Open Docker Desktop once before continuing.

macOS: Download and install Docker Desktop. Open it once to finish setup.

Linux: Follow the Docker Engine install guide for your distribution. For NVIDIA GPU support, also install the NVIDIA Container Toolkit.

Verify Docker is working by opening a terminal and running:

docker --version

You should see a version number. If you get an error, make sure Docker Desktop is open and running.

---

Step 2 — Download KTSB

Open a terminal — PowerShell on Windows, Terminal on Mac or Linux — and run:

git clone https://github.com/Dursnif/KTSB.git
cd KTSB

No git? You can also download the ZIP from GitHub and unzip it. Then open a terminal inside the unzipped folder.

---

Step 3 — Start KTSB

Linux / macOS:

cp .env.example .env
docker compose up -d

Windows (PowerShell):

copy .env.example .env
docker compose up -d

The first start downloads the Docker images (about 5 GB). This takes a few minutes depending on your internet speed. You only need to do this once — future starts are instant.

Watch the progress:

docker compose logs -f

Press Ctrl+C to stop watching logs. The services keep running in the background.

---

Step 4 — Open the browser

Go to http://localhost:5173/login in your web browser.

Log in with the default admin credentials:

Username	PIN
admin	1234

⚠️ You will be forced to change the PIN on first login. Do not skip this — the default PIN is public knowledge.

The onboarding wizard walks you through the rest — no config files to edit manually:

1. Profile — your assistant's name, language, timezone, and location (used for weather)
2. User — create your personal user account with a PIN (min. 6 digits)
3. Distribution — choose how much of the system to enable (start with Medium)
4. Integrations — connect Home Assistant, Frigate, and others (you can skip this and do it later)
5. Done — install your first AI model and start chatting

---

Connecting Home Assistant

In the KTSB admin panel, go to Settings → Home Assistant.

You need two things from your Home Assistant install:

1. The URL — the address you use to open HA in your browser. Usually something like http://homeassistant.local:8123 or http://192.168.1.x:8123.

2. A long-lived access token — this gives KTSB permission to control HA on your behalf.

To create one: in Home Assistant, click your profile picture in the bottom-left corner → Security → scroll down to Long-lived access tokens → Create token → give it a name like KTSB → copy it and save it somewhere safe.

Paste the URL and token into Settings → Home Assistant in KTSB and save. That is it.

Setting up device names — Kåre understands room names and device nicknames you define yourself. Go to Settings → Aliases to map your own names (like "the bedroom lamp") to Home Assistant entity IDs. The alias editor has a search function to help you find the right entity.

---

Connecting Frigate

In KTSB admin, go to Settings → Integrations → Frigate.

You need:

• Frigate URL — the address of your Frigate instance (e.g. http://192.168.1.x:5000)

For real-time motion event detection, also connect MQTT under Settings → MQTT with your broker address and credentials. This lets Kåre react to Frigate events as they happen.

Go to Settings → Cameras to configure which cameras Kåre should monitor, what objects to watch for (people, cars, animals), confidence thresholds, and whether to announce detections over your speakers.

---

Choosing an AI model

After the onboarding wizard, go to Settings → LLM → Manage models to pull a model. No command line needed — search by name, click Pull, and wait.

Your GPU VRAM	Recommended model	Notes
No GPU (CPU)	huihui_ai/qwen3.5-abliterated:2b	1.9 GB — slow, but works
4–6 GB	huihui_ai/qwen3.5-abliterated:4b	3.3 GB — good for light use
8–16 GB	huihui_ai/qwen3.5-abliterated:9b	6.6 GB — outperforms older 14B models
24 GB+	huihui_ai/Qwen3.6-abliterated:27b	17 GB — best results

Start by setting the same model for all roles (Kåre, Miss Kåre, Library, Mechanic) — this uses less VRAM and keeps things simple. You can assign each agent its own model later under Settings → LLM.

Why abliterated models? Kåre is a multi-agent system — agents evaluate each other, inspect logs, and ask sensitive questions internally. Standard models can refuse these internal requests, breaking the pipeline. Abliterated variants have the refusal direction removed, which makes the whole system more reliable. All recommendations above are from huihui_ai on Ollama.

Multimodal: All recommended models understand both text and images. For the full agent setup — Kåre, Miss Kåre, Library, and Mechanic — using the same multimodal model across all roles ensures camera snapshots and images can be passed between agents without gaps.

Extended thinking: Qwen3 models can pause and reason before answering complex questions. Enable it per role under Settings → LLM.

Prefer cloud? Go to Settings → LLM → Cloud and enter an API key for OpenAI, NVIDIA NIM, or another OpenAI-compatible provider.

---

Voice input

Browsers require HTTPS for microphone access. For voice input to work, Kåre needs to be served over HTTPS.

Kåre includes Caddy, which obtains a free Let's Encrypt certificate automatically. You will need a free dynamic DNS hostname from DuckDNS.

1. Create a free account at DuckDNS and register a subdomain — e.g. mykaare.duckdns.org
2. Set it to point to your home's public IP address (shown on the DuckDNS dashboard)
3. Open ports 80 and 443 in your router (search for port forwarding in your router's manual)
4. Open your .env file and set KAARE_DOMAIN=mykaare.duckdns.org
5. Restart: docker compose up -d

Voice input will work as soon as the certificate is issued — usually within 30 seconds of first start.

Access Kåre at https://mykaare.duckdns.org from any device on your local network.

For local use without voice, the default KAARE_DOMAIN=localhost works fine.

---

Service profiles

Set COMPOSE_PROFILES in your .env file to control which optional services start:

Value	What starts	When to use
(empty)	Core only (API, GUI, agents, intent)	Cloud LLM or quick test
medium	+ Semantic memory and wiki search	Recommended for most installs
full	+ Voice bridge (speech-to-text, text-to-speech)	When you want voice input and output

# .env
COMPOSE_PROFILES=medium

---

Remote access (VPN)

To access Kåre securely from outside your home network, enable the built-in WireGuard VPN service.

Requirements: a public hostname (DuckDNS recommended — you can reuse KAARE_DOMAIN) and UDP port 51820 open in your router.

Setup

1. Open port 51820 UDP in your router's port forwarding settings.
2. Generate a password hash for the VPN management web UI:

   docker run --rm ghcr.io/wg-easy/wg-easy:15 wg-easy --hash YourPassword

3. Add to your .env file:

   WG_HOST=mykaare.duckdns.org
   WG_PASSWORD_HASH=$$2a$$12$$...   # note: $$ escapes $ in Docker .env files
   

4. Add vpn to COMPOSE_PROFILES in .env:

   COMPOSE_PROFILES=medium,vpn
   

5. Restart: docker compose up -d

The VPN management UI is now available at http://your-machine-ip:51821. Add a client there, download its config or scan the QR code, and connect from your phone or laptop.

Once connected via VPN, Kåre detects you automatically as a VPN user.

Per-user access levels

In Admin → Users, set each user's remote access level:

Level	What it allows
Local only	Access only from your home network
AI chat only	Full chat and memory over VPN, no smart home control
Full access	All features available remotely

New users default to Full access. Restrict individual users under Admin → Users if needed.

Note: The VPN service requires NET_ADMIN and SYS_MODULE kernel capabilities. These work on standard Linux hosts. Some cloud VPS providers restrict kernel modules — check your host's documentation if wg-easy fails to start.

---

Updating

docker compose pull
docker compose up -d

Your configuration, state, and data are stored in folders outside the Docker images and survive updates unchanged.

---

Your data stays home

Everything in KTSB runs on your own hardware by default. No data leaves your network unless you explicitly configure a cloud LLM provider. Conversations, memories, camera snapshots, and user data stay on your machine.

---

For developers

Code-mounted mode — edit files on disk without rebuilding images:

docker compose -f docker-compose.yml -f docker-compose.dev.yml up -d

Edit a Python file → docker compose restart kaare-api (~5 seconds) → done.

For a full breakdown of every feature, agent, tool, and service profile: docs/capabilities.md

---

Project layout

configs/        ← All configuration (YAML + .env secrets)
state/          ← Runtime state (memory, timers, Qdrant snapshots)
data/           ← SQLite databases (users, long-term memory)
logs/           ← Service logs
kaare_core/     ← Shared Python library
services/       ← Embedding, voice, agents
frontend/       ← React/Vite GUI (TypeScript)

---

License

MIT