Always-on, local-only voice daemon for macOS. Hears your wake word (“clawd” by default), transcribes with whisper.cpp, then fires a configurable hook (user-defined, e.g., warelay heartbeat). Written in Go; ships with a daemon lifecycle, status socket, and launchd helper.

• Requirements: Go 1.25+, brew install portaudio pkg-config, a whisper.cpp model.
• One-liner: pnpm brabble setup && pnpm start (downloads medium Q5_1, writes config, starts daemon).
• Foreground run: go run ./cmd/brabble serve (mic + PortAudio required).

• start | stop | restart — daemon lifecycle (PID + UNIX socket).
• status [--json] — uptime + last transcripts; tail-log shows recent logs.
• mic list|set [--index N] — enumerate or select microphone (aliases: mics, microphone).
• models list|download|set — manage whisper.cpp models under ~/Library/Application Support/brabble/models.
• setup — download default model and update config; doctor — check deps/model/hook/portaudio.
• test-hook "text" — invoke hook manually; health — ping daemon; service install|uninstall|status — launchd helper (prints kickstart/bootout commands).
• transcribe <wav> — run whisper on a WAV file; add --hook to send it through your configured hook (respects wake/min_chars unless --no-wake).
• Hidden internal: serve runs the foreground daemon (used by start/launchd).
• --metrics-addr enables Prometheus text endpoint; --no-wake bypasses wake word.

• pnpm brabble — build then start daemon (default); extra args pass through, e.g. pnpm brabble --help, pnpm brabble status.
• pnpm start|stop|restart — lifecycle wrappers.
• pnpm build — build to bin/brabble; pnpm lint — golangci-lint run; pnpm format — gofmt -w .; pnpm test — go test ./....
• Lint deps: brew install golangci-lint; CI runs gofmt+golangci-lint+tests (see .github/workflows/ci.yml).

• Transcribe without the daemon: pnpm brabble transcribe samples/clip.wav
• Send through your hook (wake+min_chars enforced): pnpm brabble transcribe samples/clip.wav --hook
• Ignore wake gating for a file: pnpm brabble transcribe samples/clip.wav --hook --no-wake
• Input: any WAV; we downmix to mono and resample to 16 kHz internally.

[audio]
device_name = ""
device_index = -1
sample_rate = 16000
channels = 1
frame_ms = 20          # 10/20/30 only

[vad]
enabled = true
silence_ms = 1000      # end-of-speech detector
aggressiveness = 2
energy_threshold = -35.0  # dBFS gate; raise (e.g., -30) to suppress low-noise hallucinations
min_speech_ms = 300
max_segment_ms = 10000
partial_flush_ms = 4000  # emit partial segments (not sent to hook)

[asr]
 model_path = "~/Library/Application Support/brabble/models/ggml-large-v3-turbo-q8_0.bin"
language = "auto"
compute_type = "q5_1"
device = "auto"       # auto/metal/cpu

[wake]
enabled = true
word = "clawd"
aliases = ["claude"]
sensitivity = 0.6

[hook]
command = ""                       # REQUIRED: set to your warelay binary path
args = []                          # e.g., ["heartbeat", "--message"]
prefix = "Voice brabble from ${hostname}: "
cooldown_sec = 1
min_chars = 24
max_latency_ms = 5000
queue_size = 16
timeout_sec = 30
redact_pii = false
env = {}

[logging]
level = "info"   # debug|info|warn|error
format = "text"  # text|json
stdout = false   # also log to stdout (defaults to file-only)

[daemon]
stop_timeout_sec = 5     # wait for PID to clear on restart

[metrics]
enabled = false
addr = "127.0.0.1:9317"

[transcripts]
enabled = true

State & logs: ~/Library/Application Support/brabble/ (pid, socket, logs, transcripts, models).

• Registry: ggml-small-q5_1.bin, ggml-medium-q5_1.bin (default), ggml-large-v3-q5_0.bin.
• brabble models download <name> fetches to the models dir; brabble models set <name|path> updates config.
• brabble setup fetches the default model and writes asr.model_path; reruns doctor afterward.

• PortAudio capture → WebRTC VAD → partial segments every partial_flush_ms (suppressed from hook) → final segment; retries device open on failure.
• Wake word (case-insensitive) is stripped before dispatch; disable with --no-wake or BRABBLE_WAKE_ENABLED=0. If wake word is “clawd”, “Claude” is also accepted.
• Partial transcripts are logged with Partial=true and skipped by the hook; full segments respect hook.min_chars and cooldown.

• Default hook: ../warelay send "<prefix><text>", prefix includes hostname.
• Extra env: BRABBLE_TEXT, BRABBLE_PREFIX plus any hook.env; redaction toggle masks obvious emails/phones.
• Queue + timeout + cooldown prevent flooding; test-hook is the dry-run.

• brabble service install --env KEY=VAL writes ~/Library/LaunchAgents/com.brabble.agent.plist and prints:
  • launchctl load -w <plist>
  • launchctl kickstart gui/$(id -u)/com.brabble.agent
  • launchctl bootout gui/$(id -u)/com.brabble.agent
• service status reports whether the plist exists; service uninstall removes the plist file.

BRABBLE_WAKE_ENABLED, BRABBLE_METRICS_ADDR, BRABBLE_LOG_LEVEL, BRABBLE_LOG_FORMAT, BRABBLE_TRANSCRIPTS_ENABLED, BRABBLE_REDACT_PII (1/0).

• WebRTC VAD ships by default. Silero VAD (onnxruntime) remains an optional future path; onnxruntime is the runtime library for ONNX models and would be pulled in only if we add Silero.

• Go style: gofmt tabs (default). golangci-lint config lives at .golangci.yml.
• Tests: go test ./... plus config/env/hook coverage.
• Build: build whisper.cpp once, then:

  # headers + libs placed in /usr/local/{include,lib}/whisper (see docs/spec.md)
  export CGO_CFLAGS='-I/usr/local/include/whisper'
  export CGO_LDFLAGS='-L/usr/local/lib/whisper'
  go build -o bin/brabble ./cmd/brabble
  install_name_tool -add_rpath /usr/local/lib/whisper bin/brabble

• Models: defaults to ggml-large-v3-turbo-q8_0.bin; best quality ggml-large-v3-turbo.bin; lighter option ggml-medium-q5_1.bin. Use brabble models download <name> then brabble models set <name>.
• CI: GitHub Actions (.github/workflows/ci.yml) runs gofmt check, golangci-lint, and go test.

🎙️ Brabble. Make it say. [hooks]