noisy-coding

Voice coding for Claude Code, Jarvis-style — talk to your agent while it works. It's your voice that's noisy, not your code. Claude speaks short summaries aloud, and an always-on listener daemon turns your speech into messages Claude receives while it works — powered by the Grok (xAI) Voice API and a live "tactical HUD" dashboard.

Quick start (Claude Code plugin — any OS, nothing to clone)

The backend ships as a hardware-free Docker image (noisy/noisy-coding): all audio flows through the dashboard page (the browser tab is the microphone and the speaker). The host needs Docker and a browser — no Python, no git, no environment variables.

# terminal: install the plugin (marketplace + plugin in one line)
claude plugin marketplace add noisy/noisy-coding && claude plugin install noisy-coding@noisy

# inside Claude Code (new session):
/noisy-coding:setup

The setup command starts the published image and walks you through first contact; the plugin itself carries the MCP connection and the voice hooks (they run inside the container via docker exec, silent until the container exists). Then finish in the browser at http://127.0.0.1:8765: paste your xAI API key (console.x.ai) and click the amber ENABLE TAB AUDIO banner — that one click makes the tab your microphone and speaker. Keep the tab open and just talk.

Prefer staying inside Claude Code? Same thing, four commands: /plugin marketplace add noisy/noisy-coding → /plugin install noisy-coding@noisy → /reload-plugins → /noisy-coding:setup.

Plain Docker (no plugin, still nothing to clone)

# 1. the whole backend in one box, straight from Docker Hub
docker run -d --name noisy-coding \
  -p 127.0.0.1:8765-8767:8765-8767 \
  -v noisy-coding-config:/root/.config/noisy-coding \
  --restart unless-stopped \
  noisy/noisy-coding:latest

# 2. dashboard: API key + THIS BROWSER TAB (mic and output) in Settings
open http://127.0.0.1:8765

# 3. connect Claude Code to the MCP server in the container
claude mcp add --transport http --scope user noisy-coding http://127.0.0.1:8767/mcp

# 4. register the hooks (this is how Claude HEARS you), restart Claude Code
docker run --rm -v ~/.claude:/root/.claude noisy/noisy-coding \
  python3 /app/hooks/install.py --docker

Voice, speed, personality, language, push-to-talk, transcription mode: everything lives in the dashboard and persists across restarts (in a Docker volume). No environment variables, no config files.

Notes:

• The dashboard tab is the audio device — keep it open while talking. Speech that arrives with no tab open parks as UNHEARD; the CATCH UP button replays it when you return.
• Remote host? getUserMedia needs a secure context, so tunnel instead of exposing plain HTTP: ssh -L 8765:localhost:8765 -L 8766:localhost:8766 -L 8767:localhost:8767 host
• Linux can alternatively pass the host's PulseAudio/PipeWire socket through for native audio — see the commented variant in docker-compose.yml.

How it works (fat daemon / thin server)

All speech logic lives in one listener daemon — the single owner of the microphone, the playback queue (one voice at a time, never talking over you) and the speakers. The MCP server is a thin messenger that forwards speak requests over HTTP. Claude Code hooks deliver your transcribed speech back into the session.

mic (hardware or browser tab via WS :8766)
  -> VAD -> Grok STT -> transcript queue -> HTTP :8765
                              ^ polled by Claude Code hooks
speak (MCP, stdio or HTTP :8767) -> POST /speak -> daemon queue
  -> Grok TTS -> speakers (hardware or browser tab)

Tools

Tool	What it does
speak(text, interrupt?)	Queues text for speech and waits until it has played. Voice/speed/language come from the daemon (dashboard character), not the call.
announce(text)	Fire-and-forget variant: returns immediately, plays in the background.
change_voice(voice_id)	Deliberately switches this agent's voice (persists, shows on the dashboard).
list_voices()	Lists the built-in Grok voices (ara, eve, leo, rex, …).

Hooks (how Claude hears you)

python3 hooks/install.py registers them in ~/.claude/settings.json (user scope, idempotent — run it again after moving the repo). They are stdlib-only and run on the system python3 (3.9+): PostToolUse delivers speech while Claude works, Stop wakes an idle session when you speak, UserPromptSubmit/PreToolUse feed the dashboard's live-activity line. They fail open — with the daemon down they exit silently, so keyboard-only sessions are unaffected.

Native install (alternative to Docker)

For always-on voice without a browser tab (hardware mic/speakers), run the daemon natively. Requires Python 3.13 and uv; the API key is still configured in the dashboard, never in the shell.

git clone https://github.com/noisy/noisy-coding && cd noisy-coding
uv sync
uv run noisy-coding-listener          # first run triggers the mic prompt
claude mcp add noisy-coding --scope user \
  -- uv run --project "$PWD" noisy-coding-mcp
python3 hooks/install.py
open http://127.0.0.1:8765            # paste your xAI API key

macOS plays through the built-in afplay; installing mpv (optional) enables lower-latency streaming playback. Linux needs sudo apt install libportaudio2 mpv.

Dashboard

Everything is controlled from http://127.0.0.1:8765 (source in dashboard/, Vue 3; the legacy dashboard remains at /legacy):

• live conversation log (replay ▶ / stop ⏹ / recall ✕ on queued messages),
• real-time oscilloscope + spectrum fed by the actual mic level,
• big MUTE MIC and MUTE CLAUDE (parks speech as UNHEARD, CATCH UP replays),
• push-to-talk (hold the button or the space bar) vs automatic VAD turns,
• per-agent character (voice, speed, personality gauges),
• device pickers — including THIS BROWSER TAB as mic and output,
• costs, latencies, session ring, API-key settings.

Configuration

There is deliberately no required configuration outside the UI — the dashboard writes everything to ~/.config/noisy-coding/ (a named volume in Docker). The environment variables below exist for development and unusual setups only:

Variable	Default	Meaning
NOISY_CODING_LISTENER_PORT	8765	Port of the daemon's HTTP API
NOISY_CODING_BIND	127.0.0.1	HTTP/WS bind address (0.0.0.0 in Docker)
NOISY_CODING_STT_LANGUAGE	auto	Initial language hint (the UI selector overrides)
NOISY_CODING_MODE	batch	Initial STT mode (batch/live)
NOISY_CODING_STOP_WAIT_SECONDS	30	How long the Stop hook waits for speech
NOISY_CODING_NO_AUTOSPAWN	—	Don't auto-start the daemon from the server
NOISY_CODING_INPUT_DEVICE	system default	Initial mic (browser = the dashboard tab)
NOISY_CODING_OUTPUT_DEVICE	system	Initial speaker (browser = the dashboard tab)
NOISY_CODING_MCP_TRANSPORT	stdio	http exposes the MCP server (Docker)
NOISY_CODING_MCP_PORT	8767	MCP HTTP port (with http transport)

Development

uv run pytest                      # python tests (offline, API mocked)
cd dashboard && npm test           # frontend tests (Vitest)
cd dashboard && npm run storybook  # component workbench
cd dashboard && npm run build      # the daemon serves dashboard/dist at /

Live smoke test (spends API credits, plays audio):

uv run python scripts/smoke_test.py "Hello from Grok"