squawk
ashrocket/squawk
Voices for your Claude Code agents on macOS — fully local STT/TTS, one audio channel, radio-operator rules
Instalação
npx convez-ai ashrocket/squawk1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
squawk
Voices for your Claude Code agents on macOS. Talk to Claude hands-free, and let every agent in your cmux setup speak to you — one at a time, each with its own voice.
Fully local ears and voices: whisper.cpp (Metal) for speech-to-text, macOS neural voices or Kokoro for text-to-speech, and the Claude Code CLI as the brain. No API key, nothing leaves your Mac except what you already send to Claude.
🚧 Built in public for Mac developers running multiple Claude Code agents. Watch the repo — this is moving fast.
Why
Running several Claude Code agents in cmux means status lives in N terminal tabs. squawk gives the whole crew one audio channel with rules a radio operator would recognize:
- Focus follows mic. Run
voicein as many cmux tabs as you like — each becomes a voice agent for that window's project, but only the focused window's agent listens. The rest stand by quietly, like tabs waiting for keystrokes. (Terminal focus reporting; works in Ghostty/cmux, iTerm2, xterm.) - Conversations hold the con. An active exchange keeps the mic for 75s past the last turn, even if your focus drifts — nobody steals the channel mid-thought.
- One talker. A global speech lock — agents queue, never talk over each other.
- Barge-in (experimental). Start talking over the assistant and it stops to listen.
- Relay requests. Background agents don't grab the speakers mid-conversation;
speak --relay "need a review on PR 12"queues a message the active conversation reads aloud between turns. - Distinct voices. Each agent name is auto-assigned its own voice, best
available first: your system default, then Kokoro neural, then Apple Premium.
Assignments persist in
voices.json. - Radio protocol. The assistant ends turns with "Over." Say "over and out" to sign off.
Quick start
git clone https://github.com/ashrocket/squawk && cd squawk
./setup.sh
./voice --user YourName # two-way conversation; speak when greeted
Say "goodbye", "stop listening", or "over and out" to end.
Run it per project window: cd ~/code/my-project && ~/path/to/squawk/voice —
the agent names itself after the directory, gets its own voice, and its brain
runs in that directory, so it can read the project's files when you ask about
them. Click into a window to talk to that project's agent.
Let any agent (any cmux tab, any script) report in:
./speak --as builder --announce "Tests pass, deploy is live."
Squawk.app — installer & settings panel
Prefer clicking to shell scripts? Build the developer-focused mac app (no App Store, ad-hoc signed):
./app/build_app.sh && open app/Squawk.app
- Install — the setup checks as live status rows, one-click install for anything missing (whisper.cpp, models, Python env, Kokoro).
- Voices — audition every installed voice and check the ones agents may
use; selections persist to
pool.json, whichspeakhonors. Reassign agents, and jump to System Settings to delete unused system voices (they're SIP-protected, so only Apple's UI can remove them). - Lexicon — view and teach pronunciations.
There's also a hover-to-audition web gallery: python3 gallery/build.py && open gallery/index.html — each voice picks up the poem where the last left off.
Teach pronunciations (agents learn)
TTS butchers developer words. Fix them once, every agent learns instantly:
./speak --teach "cmux=sea mux"
Or just say it mid-conversation: "pronounce cmux as sea mux." Corrections
persist in lexicon.json and apply to all voices, including Kokoro.
How it works
mic ──▶ voice detection ──▶ whisper.cpp (Metal) ──▶ claude -p --resume
│
speakers ◀── say / Kokoro ◀── lexicon ◀── speech lock ◀───┘
voice_chat.py— the conversation loop (mic owner). Energy-based voice detection with noise-floor calibration; echo-proof: it ignores the mic while any agent holds the speech lock.speak.py— the shared voice box: speech lock, voice registry, pronunciation lexicon, Kokoro synthesis. Also a CLI.- The brain is
claude -pwith session resume — conversation continuity with whatever model you choose (--model sonnetfor smarter, slower replies).
Requirements
macOS (Apple Silicon recommended), Homebrew, Python 3.10+, Claude Code CLI. ~150MB disk for whisper, +340MB optional for Kokoro. Runs comfortably on an 8GB M1.
Roadmap (building in public)
- Two-way voice loop, local STT/TTS
- Multi-agent: speech lock, voice registry,
speakCLI - Pronunciation lexicon agents learn (
--teach, or by voice) - Window-aware: one agent per cmux tab, focus follows mic, conversations hold the con
- Relay queue: background agents message the active conversation
- Barge-in v1 (experimental; bleed-aware thresholding, no echo cancellation yet)
- STT-side pronunciation learning (recognize taught words in your speech)
- Kokoro daemon: model stays resident, spawned on demand, exits after 10 idle minutes
- Talk to your current Claude Code session, not a fresh one
- Wake word / push-to-talk modes
- Demo video
MIT © Ashley Raiteri