A self-hosted chat client for Home Assistant. Talk to your house, watch it think, see what it's doing in real time. Runs entirely on your hardware — your voice, your model, your data.
- Windows 10/11 ·
Home_x.y.z_x64_en-US.msi - macOS · coming soon
First-time install on Windows shows a SmartScreen "Unrecognized app" dialog — click
More info → Run anyway. The binary isn't code-signed yet (that's planned, but an EV cert costs real money).
- A 420×720 always-on chat window in your favourite corner, resizable to whatever shape suits you.
- Live camera card at the top of the chat — collapsed row showing camera count, expands to a tab strip + live MJPEG feed of the selected camera. Slots for V-JEPA-2 activity labels.
- Real-time view of every assistant step: thinking, tool calls, device
actions (collapsed
▸ actioncards), streaming reply text. - Voice mode in the app: click the mic, talk, get a TTS reply back. Same HA pipeline as your Voice PE.
- All conversations in one feed — typed turns from the app, voice mode from the app, and Voice PE wake-word turns at your speaker all appear in the same Home window via the chat-tee SSE stream.
- Live system metrics from the AI box: GPU util, VRAM, CPU, RAM, plus TTFT + tok/s from the local LLM.
- Dark default + parchment "paper terminal" light mode.
- Conversation memory across turns. History persists between launches.
| Piece | Purpose | Notes |
|---|---|---|
| Linux box w/ NVIDIA GPU | Runs the LLM + STT + TTS stack | RTX 4090 / 6000 / A6000 (≥ 24 GB VRAM for qwen3-32b; smaller is fine) |
| Home Assistant | Your home's brain (HAOS / Yellow / Green) | Any recent HA release with HACS support |
| Home Assistant Voice PE | Optional voice satellite | Hands-free wake-word. Not required — desktop app works standalone |
| LAN or Tailscale | Connection between desktop app and AI box | Don't expose any of this to the public internet |
- Set up the AI box → see stack/README.md (one GPU host running vLLM + Kokoro + Parakeet + the metrics sidecar via Docker Compose).
- Wire Home Assistant → see docs/SETUP.md (HACS, Extended OpenAI Conversation, Voice PE pairing, the long-lived access token).
- Install the desktop app — download the .msi from the link above, launch it, paste your HA URL + token, pick a model.
- Say "turn off the kitchen lights". Watch the action card land.
┌──────────────────────────────┐
│ Home Assistant (HAOS) │
┌────────────┐ WebSocket │ + Extended OpenAI Conv │
│ Home app │ ───────────▶ │ + Voice PE integration │
│ (Tauri) │ pipeline/run │ │
└────┬───────┘ │ │ │ │
│ │ │ OpenAI │ Wyoming │
│ HTTP 8092 │ ▼ ▼ │
│ metrics │ ┌────────────┐ ┌──────────┐ │
│ │ │ vLLM │ │ Parakeet │ │
│ │ │ qwen3-27b │ │ STT │ │
│ │ └────────────┘ └──────────┘ │
│ │ ┌──────────┐ │
│ │ │ Kokoro │ │
│ │ │ TTS │ │
│ │ └──────────┘ │
│ │ ┌─────────────┐ │
▼ │ │ metrics │ │
metrics-sidecar ───────────▶│ │ sidecar │ │
(NVML + psutil + │ └─────────────┘ │
vLLM Prometheus) └──────────────────────────────┘
│ ▲
└──────────────────────────────────┘
both run on the same Linux box
The desktop client only talks to two things: HA over WebSocket (for the conversation pipeline) and the metrics sidecar over HTTP (for system telemetry). Everything else — LLM dispatch, STT, TTS, device actions — is HA's existing pipeline working unchanged.
- Setup guide — full installation walkthrough
- Runbook — operate the stack (start/stop/recover/logs)
- Architecture — how the pieces fit
- Troubleshooting — common failures
v0.1 — text chat, tool calls, device actions, live metrics, dark + paper-terminal modes, resizable window, conversation history.
v0.2 (next) — real voice mode (mic → STT → pipeline → TTS); confirmation gating in the HA agent for security-sensitive actions; macOS .dmg build.
later — auto-update, system tray / always-on-top, EV code-signing, SQLite-backed history.
Available in the input bar:
/connect <url> [<token>]— point at an HA endpoint/token <ha long-lived access token>— replace just the token/model <name>— display label for the active model/metrics <url>— override the metrics sidecar URL/clear— wipe the feed and start a fresh conversation/demo— replay the scripted design demo/about/version— version info/help— list these commands
Ctrl/⌘ + .— stop generationCtrl/⌘ + L— clear conversationCtrl/⌘ + K— focus the input fieldEsc— cancel a pending-confirm action
MIT.