Call Code

Voice-activated Discord agent ("Jarvis") that lets you manage and communicate with Claude Code and Codex sessions running on your Mac — hands-free, from your phone.

Join a Discord voice channel, say "Hey Jarvis, what's the status of the nisra agent?" and Jarvis transcribes your speech, processes it through a Codex app-server session, communicates with your running sessions via the claude-peers-mcp broker, and speaks the response back over Discord voice.

The goal: manage your development agents while mowing the lawn.

Architecture

Phone (Discord) ←→ Discord Voice Bridge ←→ Codex App-Server (brain) ←→ Peers Broker ←→ Running Sessions

• Discord Voice Bridge — thin Node.js service handling audio I/O, wake word detection, STT/TTS. Zero intelligence.
• Codex App-Server — the brain. All NLU, intent routing, peer communication, and response generation via JSON-RPC over stdio.
• Peers Broker — localhost HTTP daemon for peer discovery and messaging between Claude Code/Codex sessions.

Audio Pipeline

Discord Opus → decode → 48kHz stereo PCM → downsample → 16kHz mono → openWakeWord (continuous)
↓ wake detected
buffer audio
↓ 2s silence
write WAV → Swift STT → text → Codex → TTS → Discord

Key Design Decisions

• macOS-native audio — SFSpeechRecognizer for STT and say for TTS. No cloud API keys for audio processing.
• openWakeWord — Python subprocess with pre-trained "hey_jarvis" model for always-on wake word detection.
• Async-first peer communication — Peer responses are polled from the broker and injected into Codex via turn/start, enabling unprompted notifications.

Prerequisites

• macOS 13+ (Ventura)
• Node.js 20+ (LTS)
• pnpm
• Python 3.9+
• Xcode Command Line Tools (for Swift compiler)
• Codex CLI installed
• claude-peers-mcp broker running
• Discord bot application (setup guide)

Setup

# Install Node dependencies
pnpm install
# Set up Python venv, build Swift STT binary
pnpm setup
# Configure environment
cp .env.example .env
# Fill in Discord credentials (see below)

Discord Bot Setup

1. Create an application at discord.com/developers
2. Go to Installation tab — set Install Link to None (required for private apps)
3. Go to Bot tab:
   • Click Reset Token and copy it
   • Enable Server Members Intent and Message Content Intent
   • Check voice permissions: Connect, Speak, Use Voice Activity
4. Invite to your server using:

   https://discord.com/oauth2/authorize?client_id=YOUR_APP_ID&scope=bot&permissions=36700160
   

5. Get IDs (enable Developer Mode in Discord Settings > Advanced):
   • Guild ID — right-click server name > Copy Server ID
   • Voice Channel ID — right-click voice channel > Copy Channel ID
   • User ID — right-click your name > Copy User ID

Environment Variables

# Discord
DISCORD_BOT_TOKEN= # bot token
DISCORD_GUILD_ID= # server ID
DISCORD_VOICE_CHANNEL_ID= # voice channel ID
DISCORD_USER_ID= # your user ID (only listens to you)# Codex
CODEX_MODEL=gpt-5.4 # model for Codex session
CODEX_CWD=/Users/you/dev # working directory for Codex# Peers Broker
PEERS_BROKER_URL=http://localhost:7899
PEERS_POLL_INTERVAL_MS=1000
# Audio
WAKE_WORD_THRESHOLD=0.5
SILENCE_THRESHOLD_RMS=500
SILENCE_DURATION_MS=2000
MAX_UTTERANCE_MS=30000
# TTS
TTS_VOICE=Evan (Enhanced) # any macOS `say` voice

List available TTS voices with say -v '?'. Enhanced/Premium voices can be downloaded in System Settings > Accessibility > Spoken Content > System Voice > Manage Voices.

Usage

pnpm dev

1. Join the configured Discord voice channel from your phone
2. Jarvis joins and speaks a greeting
3. Say "Hey Jarvis" — you'll hear a chime (listening)
4. Speak your command — after 2 seconds of silence, you'll hear a second chime (processing)
5. Jarvis processes via Codex and speaks the response

Example Commands

• "Hey Jarvis, who's online?"
• "Hey Jarvis, what's the status of the nisra agent?"
• "Hey Jarvis, tell Cedar to prioritize the downlink module."
• "Hey Jarvis, what's everyone working on?"

Async Notifications

Peer agents can send messages to Jarvis through the broker. The broker poller picks them up and Jarvis speaks them unprompted — no wake word needed for incoming messages.

Project Structure

call-code/
├── src/
│ ├── index.ts # orchestrator, state machine
│ ├── config.ts # environment loading
│ ├── discord/
│ │ ├── bot.ts # Discord client setup
│ │ └── voice.ts # voice connection, audio rx/tx
│ ├── wake/
│ │ ├── detector.ts # openWakeWord Node.js wrapper
│ │ └── detector.py # Python wake word subprocess
│ ├── audio/
│ │ ├── downsample.ts # 48kHz stereo → 16kHz mono
│ │ ├── silence.ts # RMS-based silence detection
│ │ └── wav.ts # WAV file writing
│ ├── stt/
│ │ ├── transcribe.swift # macOS SFSpeechRecognizer CLI
│ │ └── transcribe.ts # Node.js wrapper
│ ├── tts/
│ │ └── speak.ts # macOS `say` wrapper
│ ├── codex/
│ │ ├── client.ts # JSON-RPC 2.0 client
│ │ └── session.ts # Codex app-server lifecycle
│ └── broker/
│ └── poller.ts # broker message polling
├── assets/
│ ├── chime-wake.wav # listening chime
│ └── chime-processing.wav # processing chime
├── scripts/
│ ├── build-stt.sh # compile Swift binary
│ └── setup.sh # full setup script
└── python/
└── requirements.txt # openwakeword, onnxruntime

State Machine

IDLE → (wake word) → LISTENING → (2s silence) → PROCESSING → (Codex responds) → SPEAKING → IDLE
↑
(async peer message) ──┘

• Wake word during SPEAKING interrupts playback and transitions to LISTENING
• Peer messages during SPEAKING/PROCESSING are queued and delivered after

Development

pnpm dev # run with tsx --watch
pnpm typecheck # type check
pnpm build:stt # rebuild Swift STT binary