Agent Voice

sample.mp4

Local text-to-speech for people and AI agents, powered by Kokoro-82M (only English is supported).

Agent Voice creates WAV, MP3, Opus, or M4A recordings on macOS, Linux, and Windows without an API key.

Prebuilt dependencies cover macOS arm64/x64, Linux x64, and Windows x64. Linux arm64 currently needs a C build toolchain for miniaudio. Native Windows arm64 lacks an imageio-ffmpeg wheel; use x64 Python under Windows emulation.

Quick start

uv tool install agent-voice
agent-voice setup
agent-voice speak "Hello from Agent Voice." --play

agent-voice setup downloads and verifies the speech model.

How to use

# See every recording option
agent-voice speak --help

# Create a recording
agent-voice speak "The build is finished."

# Play an existing recording
agent-voice play "/absolute/path/recording.mp3"

# Create an MP3 with a readable filename
agent-voice speak "Here is your summary." --format mp3 --label summary

# Choose a voice, speed, and exact output
agent-voice speak "A slower reading." \
  --voice bf_emma --speed 0.85 --output recording.opus

# Safely pass agent text through stdin
printf '%s' "$VISIBLE_TEXT" |
  agent-voice speak --format mp3

Every agent-voice speak command prints one machine-readable JSON receipt containing the recording's absolute path, percent-encoded file_uri, audio metadata, and a delivery object with optional browser_url, audio_url, and recording_path viewer facts.

Agent Voice stores the recording plus private transcript metadata in its managed recordings directory. A lightweight localhost viewer renders the branded player document—with the recording name, native audio controls, and response text—and serves the actual WAV, MP3, Opus, or M4A recording. It starts automatically on port 8779, or on a free port when 8779 is occupied.

Agents use exactly two delivery routes:

1. Render path with the current surface's native audio player.

2. Otherwise render the installed skill's editable recording-delivery.md template using the structured receipt values:

   Agent Voice recording recording.mp3
   Listen: [web player](http://127.0.0.1:8779/player/recording.html) · [media app](file:///absolute/path/recording.mp3) · [web audio](http://127.0.0.1:8779/recordings/recording.mp3)
   ```sh
   agent-voice play "/absolute/path/recording.mp3"
   ```

The CLI owns delivery facts, not agent-facing recording prose. Each independently installable skill carries references/recording-delivery.md, where its wording can be customized. The skill derives the media link and playback command from the receipt's file_uri and path, omitting viewer links when those optional delivery facts are unavailable.

The viewer prefers port 8779 so links survive restarts. If that port is occupied, it selects a free port and reports it in the receipt. The web player renders the complete branded document, the media-app link opens the local file with the operating system default, and web audio serves the recording directly over HTTP.

The virtual player URL uses the recording name with an .html extension. If that name already belongs to another format, Agent Voice adds -2, -3, and so on.

Manage the lightweight viewer explicitly when needed:

agent-voice viewer start
agent-voice viewer stop

--output still writes the exact requested path. For HTTP delivery, Agent Voice copies that audio into the managed recordings directory instead of serving arbitrary filesystem paths or creating symlinks.

Explore the available voices and models, manage defaults, or check that Agent Voice is ready:

agent-voice voices
agent-voice models
agent-voice config --voice bf_emma --speed 1.15
agent-voice doctor --json

Defaults

Run agent-voice config to view the active persisted settings and their configuration file.

Setting	Built-in default	Save as default	Override once
Voice	af_heart	config --voice NAME	speak --voice NAME
Speed	1.0×	config --speed NUMBER	speak --speed NUMBER
Audio format	MP3	config --format FORMAT	speak --format FORMAT
Recording directory	Agent Voice's recordings/ directory	config --output-dir DIR	speak --output-dir DIR
Service	timed for 10 minutes	config --service MODE [--service-timeout MINUTES]	speak --service MODE [--service-timeout MINUTES]

on leaves the service running, off uses embedded inference, and timed stops the service after the configured number of idle minutes.

The service setting is stored as one object. Timed mode includes its duration:

{
  "service": {
    "mode": "timed",
    "timeout_minutes": 10
  }
}

Agent skills

View Agent Voice on skills.sh.

# Create speech recordings or read text aloud
npx skills add yoav0gal/agent-voice --skill create-speech-recording --global --agent codex --yes

# Add audio to requested written responses
npx skills add yoav0gal/agent-voice --skill spoken-response --global --agent codex --yes

Use --agent '*' instead of --agent codex to install the same skills for all agent destinations recognized by the skills CLI.

Local speech API

agent-voice serve

curl http://127.0.0.1:8765/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{"input":"The task is complete.","voice":"af_heart","response_format":"mp3"}' \
  --output speech.mp3

The speech API binds only to localhost. It is separate from the lightweight recording viewer. agent-voice speak uses the speech API automatically when available and falls back to embedded inference.