Vox

Local shared voice for every agent on your Mac — one mic, many voices, no cloud STT.

Vox — shared waveform with multiple agent nodes

Whisper + Kokoro local Shared session FIFO queue Menu bar

What it does

Vox is a persistent local voice runtime exposed as an MCP server. Claude, Codex, Grok, Fable — any host on the machine — can speak and listen through the same session. Each agent keeps its own voice. Hardware is serialized by a queue, not by exclusive ownership theater.

Working today

Shared session — nobody “owns” the mic; agents wait their turn
Per-agent voices via ?agent= and ~/.vox/agents.json
Talk / Narrate / Dictate IO modes from the status panel
Finished STT survives host cancel — undelivered recovery
Mic opens only for a bounded listen; never ambient recording
Loud-room endpointing so the mic doesn’t stick open forever

Stack

Python · MCP HTTP · Whisper.cpp · Kokoro · Swift menu bar · launchd

Why it exists

Agent voice was either a paid cloud pipe or a flaky exclusive lease that made multi-agent work feel broken. Vox is the local shared runtime: private, queue-fair, and boring enough to trust for real work.