claude-voice

Talk to a coding agent on your phone. Push to talk, watch the transcript, hear the reply spoken back. The agent is the claude CLI running in Termux on the same device — so it has your repos and tools — and the app is a thin voice front-end that reaches it over localhost. Nothing leaves the phone.

  

flowchart TB
  subgraph phone["📱 one phone — nothing leaves the device"]
    direction TB
    subgraph app["Android app · this repo"]
      direction TB
      mic["🎙️ hold-to-talk<br/>16 kHz mono wav"]
      tx["📝 transcript"]
      spk["🔊 text-to-speech"]
    end
    subgraph bridge["Termux · bridge/ · Go · 127.0.0.1:8765"]
      direction TB
      stt["whisper.cpp<br/>speech → text"]
      agent["claude -p --continue<br/>your repos + tools"]
    end
  end

  mic -- "POST /stt" --> stt
  stt -- "text" --> tx
  tx -- "POST /chat" --> agent
  agent -- "reply (streamed)" --> spk

  classDef appbox fill:#ede9fe,stroke:#7c3aed,color:#1e1b4b;
  classDef bridgebox fill:#ecfdf5,stroke:#059669,color:#064e3b;
  class mic,tx,spk appbox;
  class stt,agent bridgebox;

Loading

Layout

• app/ — Android app (Kotlin, minSdk 23): hold-to-talk records 16 kHz mono PCM, POSTs to the bridge, shows the transcript, speaks the reply.
• bridge/ — Termux-side localhost server, a single static Go binary: N agents (one per dir), /stt via whisper.cpp, /chat via the claude CLI.
• .drone.jsonnet — CI: builds the APK and publishes to GitHub releases on tag.

Quick start

1. In Termux, build whisper.cpp and put claude on PATH, then run the bridge from your repo: cd <repo> && ./claude-voice-bridge-arm64 (serves 127.0.0.1:8765).
2. Install the app APK (CI artifact or local build), open it, confirm the bridge URL, grant the mic permission, hold to talk.

Build

• Bridge: ./bridge/build.sh — a static GOOS=android binary at the repo root.
• App: needs the Android SDK (compileSdk 34); CI builds it (see .drone.jsonnet).

Bridge API

Method	Path	Body	Returns
GET	/health	—	ok
GET	/agents	—	[{id,name,dir,branch,dirty}]
POST	/agents	{"dir":"~/repo"}	current agent list
DELETE	/agents/<id>	—	current agent list
GET	/ls?dir=	—	{dir,parent,dirs}
POST	/stt	WAV bytes	transcript (whisper.cpp)
POST	/chat	{"text","agent":id}	agent reply, streamed (claude -p)
GET	/voices	—	Piper voices ([] if off)
POST	/tts	{"text","voice"}	WAV audio (Piper) or 501 if off

One agent per directory, each with its own claude --continue conversation.

Config (env)

• VOICE_PORT / VOICE_HOST — default 8765 / 127.0.0.1
• VOICE_PERM — claude permission mode (default bypassPermissions)
• VOICE_TIMEOUT — seconds before a stuck turn aborts (default 1800)
• VOICE_WORKDIR — initial agent dir (default: home)
• WHISPER_BIN / WHISPER_MODEL, and the PIPER_* paths

Piper voices (optional)

Install with ./install-piper.sh (needs glibc-runner). When piper is present at ~/piper, the bridge serves /tts and /voices; otherwise the app falls back to Android TTS. Add voices by dropping <name>.onnx[.json] into ~/piper-voices.