Give your AI assistant a real phone.

Got a spare rootable Android and a spare phone line? Bandophone lets your AI assistant make and receive actual phone calls — dial real numbers, talk to real people, handle real conversations.

No SIP infrastructure. No number porting. No dedicated hardware. Just a phone that your AI can use like a person would.

⚠️ Active Development — Core audio pipeline working. Clawdbot integration in progress.

🛠️ Built by a UX designer and an AI — This is a best-effort project by Noah (UX designer, not a systems programmer) and Bando (an AI assistant). We're learning as we go. Do not assume this is secure. Don't use it for anything sensitive. We're sharing it because the concept is cool and maybe real developers can help make it better.

Your AI assistant shouldn't be trapped in a chat window. It should be able to:

• 📞 Call businesses — Schedule appointments, check hours, make reservations
• 📲 Receive calls — Answer your phone when you're busy, take messages
• 🗣️ Have real conversations — Using OpenAI Realtime API with sub-second latency

All over the regular phone network, with a real phone number that anyone can call.

┌─────────────┐     ┌─────────────────────────────────────────┐     ┌─────────────┐
│   Cellular  │────▶│            Bandophone App               │────▶│  Clawdbot   │
│    Call     │◀────│  TinyALSA → OpenAI Realtime → TinyALSA  │◀────│  Gateway    │
└─────────────┘     └─────────────────────────────────────────┘     └─────────────┘
                                       │
                                       ▼
                              • Context (memories, calendar)
                              • Transcript logging
                              • Tool execution (lights, reminders)

1. Capture: TinyALSA captures far-end audio from device 20 (48kHz stereo)
2. Convert: Downsample to 24kHz mono for Realtime API
3. AI Processing: Stream to OpenAI with server-side VAD and barging
4. Playback: TinyALSA injects AI audio via device 19 into the call

1. Context Injection: On call start, fetch memories, calendar, user info
2. Transcript Logging: Stream conversation to daily memory files
3. Tool Bridging: AI can check calendar, create reminders, control lights

• Rooted Android phone (Magisk recommended)
• Pixel 7 Pro (other devices need adaptation)
• SELinux permissive or appropriate policies
• TinyALSA binaries on device (/data/local/tmp/)
• OpenAI API key with Realtime API access
• Clawdbot Gateway (optional, for full integration)

• Map PCM devices for call audio on Pixel 7 Pro
• Identify mixer controls for far-end capture
• Document audio routing (48kHz stereo native)
• Verify TinyALSA capture/playback during calls

• TinyALSAStreamer: Process-based capture/playback
• Sample rate conversion: 48kHz ↔ 24kHz
• Stereo ↔ mono conversion
• FIFO-based continuous playback

• HybridRealtimeBridge: Full Realtime API integration
• Server-side VAD (voice activity detection)
• Barging (interrupt AI while speaking)
• Whisper transcription of both parties
• Function calling support

• ClawdbotBridge: RPC client for Gateway
• Context fetching (memories, user info, calendar)
• Transcript streaming
• phone-bridge plugin for Gateway
• Tool execution (calendar, reminders, lights)
• Full testing with live calls

• Call detection and auto-answer
• Background service
• API key secure storage
• Settings UI

bandophone/
├── android/
│   └── app/src/main/java/com/bando/phone/
│       ├── audio/
│       │   ├── HybridRealtimeBridge.kt  # Main orchestrator
│       │   └── TinyALSAStreamer.kt      # TinyALSA wrapper
│       └── bridge/
│           └── ClawdbotBridge.kt        # Gateway RPC client
├── docs/
│   ├── CLAWDBOT_PHONE_CHANNEL_PLAN.md   # Integration design
│   ├── HYBRID_REALTIME_AUDIO_PLAN.md    # Audio architecture
│   ├── AUDIO_ROUTING.md                 # Device findings
│   └── NATIVE_AUDIO_ARCHITECTURE.md     # TinyALSA details
└── working-docs/                        # Development notes

# Push TinyALSA binaries
adb push tinycap tinyplay tinymix /data/local/tmp/
adb shell chmod +x /data/local/tmp/tiny*

# Test capture (during active call)
adb shell su -c "/data/local/tmp/tinycap /data/local/tmp/test.wav -D 0 -d 20 -c 2 -r 48000 -b 16 -p 480 -n 4"

Add to your Clawdbot config:

extensions:
  phone-bridge:
    enabled: true
    authToken: "your-secret-token"
    trustedNumbers:
      - "+16305382264"  # Your number
    logTranscripts: true

cd android
./gradlew assembleDebug
adb install app/build/outputs/apk/debug/app-debug.apk

Device 20 (Capture):  Far-end audio (what caller says)
Device 19 (Playback): Inject audio (what AI says)
Mixer 167: Mic mute control
Mixer 152: Capture routing (set to DL for far-end)

Supported OpenAI Realtime voices:

• alloy - Neutral, balanced (default)
• echo - Warm, conversational male
• shimmer - Clear, expressive female
• ash - Soft, thoughtful
• ballad - Warm, storytelling
• coral - Bright, friendly
• sage - Calm, wise
• verse - Dynamic, engaging

This is uncharted territory. If you have:

• A rooted Android device and want to help test
• Knowledge of Android audio internals
• Experience with real-time audio streaming

Please open an issue or PR! Device-specific findings are especially valuable.

MIT

• Built for the Clawdbot community
• Designed to integrate with OpenAI Realtime API
• Inspired by the dream of AI assistants that can actually call people

• Clawdbot — The AI assistant framework this was built for
• Clawdbot Android Node — Companion app

"The best interface is no interface." — Golden Krishna

Sometimes the best interface is just a phone call.