setup

Set up the noisy-coding voice backend for this user. Follow these steps in order and narrate briefly as you go:

1. Check Docker: run docker --version. If missing, tell the user to install Docker Desktop (or the docker engine on Linux) and stop.
2. Check whether the backend already runs: curl -s -m 2 http://127.0.0.1:8765/status. If it answers, skip to step 4.
3. Start the backend from the published image (multi-arch, hardware-free — all audio flows through the dashboard page). ALWAYS pull first — docker run silently reuses a stale local latest, and a version-skewed container breaks the hooks in ways that are miserable to debug:

   docker pull noisy/noisy-coding:latest
   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
   

   If the container name already exists, still pull; when the pull brought a new image, recreate the container (docker rm -f noisy-coding, then the docker run above — the config volume preserves the API key); otherwise just docker start noisy-coding.
4. Verify the services NON-INVASIVELY — exactly like this, no raw TCP probes:
   • curl -s http://127.0.0.1:8765/status must return JSON (the daemon),
   • docker port noisy-coding must list 8765, 8766 and 8767 (do NOT nc -z port 8766 — a bare TCP open on the WebSocket bridge dumps harmless-but-scary handshake tracebacks into the container logs),
   • the MCP endpoint needs no probe; if you must, know that a plain curl http://127.0.0.1:8767/mcp returning 406 Not Acceptable IS the healthy answer (MCP requires an Accept: text/event-stream header).
5. Tell the user to finish in the browser at http://127.0.0.1:8765 —
   • paste their xAI API key when asked (from console.x.ai),
   • click the big amber ENABLE TAB AUDIO banner (the container preselects this tab as microphone and speaker; the click grants the mic permission),
   • keep that tab open while talking.
6. CLOSE THE LOOP — do not declare success yet. Ask the user to say when they've pasted the key and clicked the banner, then poll curl -s http://127.0.0.1:8765/status and confirm BOTH "api_key_set": true AND "tab_audio": true. Only then report the setup as complete. If either stays false, debug with the user (key rejected? mic permission denied?).
7. The plugin already registered the MCP server and the voice hooks. On a first install the server usually shows as failed at this point — Claude Code tried to connect it at session start, before the container existed, and its ~30 s client-side timeout expired. NO full restart is needed: tell the user to run /mcp, pick noisy-coding, and choose reconnect — the launcher respawns, finds the container running, and the speak tool appears in this same session. (A restart works too, but reconnect is one menu action.) Without the speak tool Claude cannot answer aloud, though the user's speech already comes through. After the restart, messages tagged [VOICE] are the user's real spoken words delivered by this plugin's hooks — see the plugin's voice-conversations skill for the conventions.

Do not configure anything through environment variables — every setting lives in the dashboard UI.