OpenDash

Turn any Android tablet into a voice-first smart-home device with a local AI brain.

Think "Alexa without the cloud, OpenClaw-level autonomy, all on-device."

Say a wake word, ask anything, and the agent responds with voice, controls smart-home devices, runs timers, fetches weather, reads your notifications, or follows user-taught skills — all without a network call unless you explicitly ask for one.

---

Why this project

Existing smart speakers force you to choose between:

• Cloud assistants (Alexa / Google Home / Siri) — great UX but record your voice, lock you into ecosystems, and die the moment the vendor's servers do.
• Power-user setups (Home Assistant + ESPHome + Whisper + Piper) — private, but require stitching 5 services together to match basic Alexa features.
• Agent frameworks (OpenClaw, Hermes) — powerful on a laptop, unusable on a kitchen countertop tablet.

OpenDash is all three at once, on a single Android tablet you already own.

---

What it does today

Smart-home device ("Alexa feel")

• Custom wake word (Vosk) with always-on foreground service
• Sub-200ms fast-path for 40+ common intents (see full catalog below) — lights, volume, timers, alarms, weather, forecast, calendar, notifications, location, find-device, briefings, routines, app launch (fuzzy), media control, presence ("I'm home" / "I'm leaving"), goodnight, device health, battery, lock screen, open URL, Settings deep-links, multi-room broadcast + handoff
• Voice-first UI: breathing VoiceOrb shows pipeline state (listening / thinking / speaking) with audio-level reactive scaling
• Error recovery copy: "I didn't catch that. Try again?" not "list index out of range"
• Filler phrases ("Got it", "One moment") while the LLM is working
• Audio focus handling, barge-in, continuous conversation mode
• Battery-saver + thermal-throttle aware; model-download resume over flaky Wi-Fi

Fast-path catalog (click to expand — 40+ matchers)

Utterances handled directly by FastPathRouter without round-tripping through the LLM (<200ms target). All matchers are EN + JA aware. Ordering in DEFAULT_MATCHERS is significant.

Matcher	Sample utterances	Tool
BroadcastCancelTimerMatcher	"cancel timers on all speakers", "全スピーカーのタイマーをキャンセル"	broadcast_cancel_timer
CancelAllTimersMatcher	"cancel all timers", "タイマー全部止めて"	cancel_all_timers
BroadcastTimerMatcher	"set a 5 minute timer on all speakers", "全スピーカーで5分タイマー"	broadcast_timer
TimerMatcher	"set timer for 5 minutes", "5分タイマー"	set_timer
AlarmMatcher	"wake me up at 6:30", "7時にアラーム"	set_timer (wall-clock)
TimeQueryMatcher	"what time is it", "今何時"	get_datetime
VolumeMatcher	"volume up/down", "set volume to 50", "mute", "ミュート"	set_volume
ThermostatMatcher	"set thermostat to 22", "エアコン25度"	execute_command climate
FanMatcher	"fan on/off", "扇風機を消して"	execute_command fan
TvMatcher	"TV on/off", "テレビをつけて"	execute_command media_player
LockMatcher	"lock/unlock the door", "ドアをロック"	execute_command lock
CoverMatcher	"open the blinds", "カーテンを開けて"	execute_command cover
EverythingOffMatcher	"turn off everything", "全部消して"	execute_command light off
LightsMatcher	"lights on", "dim the lights", "明るさ80%", "bedroom lights off"	execute_command light
MediaControlMatcher	"pause music", "next track", "再生して"	execute_command media_player
RunRoutineMatcher	"run X routine", "Xルーチンを実行"	run_routine
SettingsMatcher	"open wifi settings", "Wi-Fiの設定", "アクセシビリティ"	open_settings_page
OpenUrlMatcher	"open https://example.com", "go to github.com"	open_url
LaunchAppMatcher	"open camera", "Chromeを開いて", "天気アプリ開いて"	launch_app
FindDeviceMatcher	"find my tablet", "デバイスを探して"	find_device
GoodnightMatcher	"goodnight", "おやすみ"	goodnight (composite)
ArriveHomeMatcher	"I'm home", "ただいま"	arrive_home (composite)
LeaveHomeMatcher	"I'm leaving", "行ってきます"	leave_home (composite)
MorningBriefingMatcher	"morning briefing", "朝のサマリー"	morning_briefing
EveningBriefingMatcher	"evening briefing", "夜のサマリー"	evening_briefing
ForecastMatcher	"weather tomorrow", "明日の天気"	get_forecast
WeatherMatcher	"what's the weather", "今日の天気"	get_weather
NewsMatcher	"news", "ニュース"	get_news
CalendarMatcher	"what's on my calendar today", "今日の予定"	get_calendar_events
ClearNotificationsMatcher	"clear notifications", "通知を消して"	clear_notifications
ListNotificationsMatcher	"show notifications", "通知一覧"	list_notifications
LocationMatcher	"where am I", "現在地を教えて"	get_location
ListMemoryMatcher	"what do you remember", "覚えていること"	list_memory
ListDevicesMatcher	"list my devices", "デバイス一覧"	get_devices_by_type
ListTimersMatcher	"list timers", "タイマー一覧"	get_timers
DeviceHealthMatcher	"system status", "storage space", "システム状態", "ストレージ残量"	get_device_health
BatteryMatcher	"battery level", "バッテリー残量"	BatteryMonitor (speak-only)
LockScreenMatcher	"lock the tablet", "画面をロック"	lock_screen
ListPeersMatcher	"list nearby speakers", "近くのスピーカー"	list_peers
BroadcastGroupMatcher	"broadcast X to kitchen", "キッチンに X ってアナウンス"	broadcast_tts (group)
BroadcastTtsMatcher	"tell all speakers X", "全スピーカーにアナウンス：X"	broadcast_tts
HandoffMatcher	"move this to the kitchen speaker", "キッチンにハンドオフ"	handoff_session
FlipCoinMatcher	"flip a coin", "コインを投げて"	flip_coin
DatetimeMatcher	"what's today's date", "今日は何日"	get_datetime
GreetingMatcher	"thanks", "ありがとう", "おはよう"	speak-only
HelpMatcher	"help", "what can you do", "できることを教えて"	speak-only

Full notes with precedence rules in docs/fast-paths.md.

Local AI agent ("OpenClaw feel")

• On-device LLM via LiteRT-LM (Gemma 3n / 4 family) with GPU→CPU fallback and hardware-tier-tuned context size / thread count

• Pluggable chat templates (Gemma / Qwen / Llama3 / ChatML)

• ReAct agent loop with multi-tool chaining (AgentPlan), up to 10 tool rounds

• Context compaction keeps long conversations coherent

• Device state auto-injected into every prompt (<device_state>)

• Skills system — drop a SKILL.md into assets/skills/<name>/ or install one at runtime via install_skill_from_url (OpenClaw-style). 36 bundled skills:

  Bundled skills catalog (click to expand)

  Skill	Purpose
  arrival-home	Warm welcome on arrival — entry lights, next event, urgent notifications
  bedtime-for-kids	Family wind-down — broadcast warning, dim lights, pause media, cool-down timer
  bedtime-routine	Evening wind-down — pause media, dim lights, set alarm, summary
  breathing-exercise	4-7-8 breathing guide — four voice-cued cycles
  cooking-assistant	Recipes, steps, unit conversions, kitchen timers
  cooking-session	Multi-timer cooking — labeled parallel timers + "dinner is ready" broadcast
  dinner-call	Broadcast TTS to every speaker + pinned banner
  dog-walk	Weather cue, 30-min backstop timer, welcome-home prompt
  eye-break	20-20-20 screen-break reminder
  focus-mode	Deep work — quiet alerts, steady light, optional Pomodoro
  gratitude-journal	Nightly 3-item gratitude prompt stored via remember
  guest-mode	Relax the agent for visitors — silence wake-word + personal data
  home-control	Smart-home control via HA and other providers
  hydration-reminder	Low-friction hourly water nudge
  leaving-home	Send-off — lights out, battery reminder, weather preview
  meditation	Quiet session — dim warm light, silence notifications, optional bell
  morning-routine	Wake-up flow — briefing, lights, ambient music, schedule
  movie-night	Dim lights, pause notifications, queue media
  news-briefing	Daily news + weather briefing
  party-mode	Colourful lights, music, volume up, multi-speaker announcements
  pomodoro	25/5 cycles with 15-min rest every fourth block
  power-nap	20-min nap — dim to 20%, short timer, gentle wake
  quick-note	Rapid voice-note capture into notes memory namespace
  quiet-hours	Night-time noise cap — low TTS, skip broadcasts
  rainy-day	Forecast-driven cozy indoor plan suggestion
  reading-time	Warm light, silence everything else, optional timer
  remind-me-tomorrow	Dated memory entry surfaced in morning-briefing
  sick-day	Rest mode — dim lights, low volume, hydration every 2h
  stretch-break	5-min guided stretch routine
  study-mode	Quiet study — dim lights, silence, optional Pomodoro
  task-manager	Capture/recall to-do items via long-term memory
  travel-mode	Trip mode — all lights off, cancel timers, goodbye broadcast
  voice-assistant	General voice interaction rules (loaded every turn)
  wake-up-gently	Sunrise-style 10-min light ramp + briefing
  where-did-i-put-it	Search memory for prior "I put X in Y" statements
  workout	Bright lights, upbeat music, interval timers

• 50+ built-in LLM tools:

  • System: timer, volume, app launcher, datetime, notifications, calendar, contacts, SMS, camera, photos, screen record, device health, location, screen reader
  • Info: weather (Open-Meteo), search (DuckDuckGo), news (RSS), web fetch, unit converter, calculator, currency converter
  • Memory: remember, recall, search_memory, semantic_memory_search, forget
  • Documents: RAG ingest + retrieve with TF-IDF scoring
  • Routines: user-defined multi-step workflows
  • Skills: get_skill, list_skills, install_skill_from_url

Smart-home control

Protocol	Devices
Home Assistant	Everything HA exposes (lights, climate, covers, media, sensors)
SwitchBot	Bot, Curtain, Plug, Bulb, Strip/Ceiling Light, Lock, Meter
Matter	Android Matter API commissioning
MQTT	Shelly, Tasmota, and any MQTT-discoverable device

Tablet control (no root)

• Accessibility service drives: read_active_screen, tap_by_text, scroll_screen, type_text
• Notification listener drives: list_notifications, clear_notifications, reply_to_notification
• Fuzzy launch_app finds the right app even when the user's wording is loose ("open the weather app")
• Deep-link into Settings, open URLs, pin routines to the home screen as dynamic shortcuts
• Quick Settings Talk tile for one-tap voice without the wake word
• Opt-in Device Admin unlocks lock_screen; no other policies requested
• Full recipe book in docs/tablet-control-cookbook.md

Multi-room (_opendash._tcp. on port 8421)

• Auto-discover peers via mDNS; pair with a shared HMAC-SHA256 secret
• 4-word pairing fingerprint so users verify secrets match without typing them again
• Speaker groups (kitchen, upstairs, ...) so broadcast_tts can target subsets
• Tools: broadcast_tts, broadcast_timer, broadcast_cancel_timer, broadcast_announcement, handoff_session, list_peers
• Quickstart in docs/multi-room-quickstart.md, full recipe book in docs/multi-room-cookbook.md, wire format in docs/multi-room-protocol.md

Privacy / data

• All AI runs locally. The LLM lives on your tablet, not a server.
• Secrets encrypted via EncryptedSharedPreferences (AES256-GCM).
• Network calls minimized and opt-in. See SECURITY.md.

Languages

UI translated into English, 日本語, Español, Français, Deutsch, 简体中文 — pick in Settings → Language (Android 13+) or follow the system locale. Fast-path matchers and bundled skills respond to EN + JA utterances side-by-side. See docs/i18n.md for the full list and how to contribute a new locale.

---

Hardware recommendations

RAM	Model tier
3-4 GB	Gemma 3n E2B (Q4) ~1.2 GB
4-6 GB	Gemma 3n E2B / Qwen3 1.5B
6-8 GB	Gemma 3n E4B (Q4) ~2.5 GB
8+ GB	Gemma 4 / Phi-4 Mini

HardwareProfile auto-detects RAM and tunes context / GPU layers / thread count.

---

Setup (5 minutes)

1. Install the APK (GitHub Releases or ./gradlew assembleDebug)
2. Grant microphone permission on first launch
3. Model auto-downloads (~2 GB, public HuggingFace)
4. Settings → configure device providers (SwitchBot / MQTT / HA / Matter)
5. Say your wake word and start asking

Optional permissions unlock more tools:

• Calendar → get_calendar_events
• Contacts → search_contacts
• Location → get_location
• Notification access → list_notifications, clear_notifications, reply_to_notification
• Accessibility service → read_screen, read_active_screen, tap_by_text, scroll_screen, type_text
• Device Admin (opt-in) → lock_screen (only force-lock policy requested)
• SMS → send_sms (with LLM confirmation prompt)
• Photos → list_recent_photos
• Camera → take_photo

---

Development

Kotlin 2.1 + Jetpack Compose + Material 3 + Hilt. Built against SDK 35.

./gradlew testDebugUnitTest   # ~1000 unit tests
./gradlew assembleDebug       # build APK
./gradlew lint                # Android Lint (baseline at app/lint-baseline.xml)
./gradlew jacocoTestReport    # coverage report → app/build/reports/jacoco/

JUnit 5 + MockK + Truth + MockWebServer + Turbine, 80%+ coverage target for non-UI code. See CONTRIBUTING.md.

Architecture

app/src/main/java/com/opendash/app/
├── assistant/
│   ├── provider/embedded/   # LiteRT-LM on-device provider
│   ├── provider/openai/     # OpenAI-compatible REST+SSE
│   ├── provider/openclaw/   # OpenClaw WebSocket
│   ├── router/              # 4 routing policies
│   ├── skills/              # SkillRegistry + SkillInstaller (OpenClaw SKILL.md)
│   ├── routine/             # Routine + RoomRoutineStore
│   ├── agent/               # AgentPlan + PlanExecutor
│   ├── proactive/           # Time-based suggestions + rules
│   └── context/             # ContextCompactor + DeviceContextBuilder
├── device/                  # HA / SwitchBot / MQTT / Matter providers
├── voice/
│   ├── pipeline/            # VoicePipeline + ErrorClassifier
│   └── fastpath/            # FastPathRouter (Alexa-style LLM bypass)
├── tool/
│   ├── system/              # timer / volume / notifications / SMS / camera etc
│   ├── info/                # weather / search / news / calc / currency
│   ├── memory/              # persistent KV + TF-IDF semantic search
│   ├── rag/                 # document chunker + retrieval
│   ├── accessibility/       # screen reader service
│   └── analytics/           # usage stats
├── ui/                      # Chat / Dashboard / Ambient / Settings / Home
├── service/                 # Foreground + VoiceInteraction + Boot
└── data/                    # Room + encrypted preferences

See docs/architecture.md for the deeper dive.

---

Priority order

Every change in this project advances one of these, in order:

1. Smart home device feel — it has to feel like Alexa for everyday use
2. Local agent capabilities — OpenClaw-style, on-device
3. UX polish — ambient, onboarding, error recovery
4. Hybrid gateway — connect to external OpenClaw / Hermes when needed
5. Refactor / quality — code, performance, security
6. OSS project health — docs, CI, community

See docs/roadmap.md for the full per-item list.

---

Inspiration / credit

• OpenClaw — the agent framework we bring to Android
• Ava — voice state machine + wake ripple UX patterns
• off-grid-mobile-ai — hardware-aware LLM init + context compaction
• dash-voice — tablet-first smart-home UX
• local-llms-on-android — LiteRT-LM integration reference

Each of these informed a specific PR; see commit messages for "stolen from ..." notes.

---

License

MIT