dial — give your Claude Code a phone number

Talk to your Claude Code, and have it call you back when the work's done. Type as usual, or talk; Claude answers out loud in Dial's voice. After dial login, a phone call to Dial routes into your session — so you can hand off a task, hang up, and get rung back with the result. It does not modify or fork Claude Code — it runs the real claude inside a pseudo-terminal and layers voice + a bridge on top.

┌──────────────── dial (wrapper) ───────────────┐
│  your keyboard ──────────────► the real       │
│  Ctrl-T / hands-free mic ────►  `claude` (pty) │
│                                                │
│  claude reply ── Stop hook ──► daemon ── TTS ──► 🔊
│                        └── <callme> ─► Dial rings your phone 📞
│  a Dial phone call ◄──► bridge ◄──► this session
└────────────────────────────────────────────────┘

• Voice in — Ctrl-T or hands-free; your speech streams to Deepgram live STT and is typed into the same session as if you'd typed it.
• Voice out — a Stop hook sends each reply to a tiny local daemon that speaks it.
• Bridge — after dial login, a phone call to Dial reaches this session, and Claude can ring you back when a task finishes.

Requirements

• macOS (playback uses afplay)
• Node ≥ 22 (the bridge uses the built-in WebSocket)
• ffmpeg — brew install ffmpeg (microphone capture)
• DEEPGRAM_API_KEY in your environment (or a .env in the cwd)
• claude on your PATH

Setup

cd voice-cli
npm install          # builds node-pty
npm link             # makes `dial` global
dial login           # connect your phone (opens the browser, sign in)

Use

Instead of claude, run:

dial                 # shows a quick menu: new / continue / resume / hands-free
dial --resume        # or pass claude flags straight through
dial "fix the build" # a prompt goes straight to a new session

• Talk: press Ctrl-T, speak, Ctrl-T to send. (Talk when the prompt is empty.)
• Hands-free: press Ctrl-G (or launch with --handsfree, or say "go hands-free" into a Ctrl-T recording). The mic stays open — just talk; a ~1.2 s pause sends your words. The mic is muted while Dial speaks, so it never hears itself. Leave with Ctrl-G or by saying "hands-free off" / "stop listening".
• Type: just type — Claude is completely normal underneath.
• Barge-in: pressing Ctrl-T while it's talking cuts the speech off (and in push-to-talk mode starts recording). Esc just silences it.
• Replies are spoken short: the screen keeps Claude's full answer, but only a 1–2 sentence conversational summary is read aloud (Claude appends it in a <say>…</say> block; if it doesn't, we speak the first sentence or two).

Check your setup any time:

dial-voice doctor
dial-voice mic       # live mic test — prints what it hears as you speak

Config (env)

| Var | Default | What | |-----|---------|------| | DEEPGRAM_API_KEY | — | required | | DIAL_VOICE_MODEL | aura-2-… (from DEEPGRAM_TTS_MODEL) | the spoken voice | | DEEPGRAM_MODEL | nova-3 | STT model | | DIAL_VOICE_PORT | 8788 | local speak daemon | | DIAL_MIC | auto-detected | avfoundation audio-device index | | DIAL_MAX_SECONDS | 30 | max single push-to-talk utterance | | DIAL_PTT | 20 (Ctrl-T) | push-to-talk key byte | | DIAL_HF_KEY | 7 (Ctrl-G) | hands-free toggle key byte |

Find your mic index if 0 is wrong:

ffmpeg -f avfoundation -list_devices true -i "" 2>&1 | grep -A20 audio

Phone bridge (call your Claude from anywhere)

With the bridge configured, this session registers with the Dial server as your "local Claude." On a Dial call (iOS app / phone), say "talk to my Claude" — the server routes your words here as text, Claude answers, and the reply is spoken into the call. Audio never leaves the call; your Mac exchanges text only.

DIAL_USER_ID=<your dial user id> DIAL_BRIDGE_SECRET=<server DIAL_INTERNAL_SECRET> dial-voice

(or put both in your .env). dial-voice doctor shows the bridge status. While a phone turn is pending, the reply goes to the call instead of the local speakers.

Notes

• First run prompts for microphone permission (macOS gives it to your terminal app).
• dial-voice init installs the hooks globally instead (opt-in); the default never touches ~/.claude/settings.json.
• Set DIAL_VOICE_LOG=/tmp/hook.log to debug the speak hook.