agent-voice 0.10.0

Agent Voice gives AI agents the ability to create local speech recordings.

AI agents produce useful work, but they still communicate mostly through text. Every response competes for a developer's visual attention, so valuable work is often skimmed or missed.

My initial use case was listening alongside the text: text-to-speech for coding agents such as Codex and Claude Code.

The larger idea is simple: audio should be a first-class medium for agents.

Agent Voice runs locally, uses Kokoro-82M, and requires no API key. English is the best-supported language. The bundled 54-voice catalog also covers Japanese, Mandarin Chinese, Spanish, French, Hindi, Italian, and Brazilian Portuguese, though quality varies.

Listen to the voice catalog.

Install and setup

Install the CLI with uv:

uv tool install agent-voice

Or with pipx:

pipx install agent-voice

Then download the speech model:

agent-voice setup

For setup with terminal playback controls, see Terminal playback controls.

Test the CLI directly:

agent-voice speak "Hello from Agent Voice." -p

Choose a skill

Agent Voice provides two skill types:

• create-speech-recording turns supplied text into audio. Use it to create a recording or read something aloud.
• spoken-response creates the spoken semantic twin of an assistant response. It can speak the current response, the previous response, or later responses in the thread.

Choose the variant based on where you run your agent

Normal skills (most portable)

npx skills add yoav0gal/agent-voice -g --skill create-speech-recording
npx skills add yoav0gal/agent-voice -g --skill spoken-response

Desktop skills (best experience for desktop app users)

npx skills add yoav0gal/agent-voice -g --skill create-speech-recording-desktop
npx skills add yoav0gal/agent-voice -g --skill spoken-response-desktop

Controls skills (best experience in supported terminals)

agent-voice controls install
npx skills add yoav0gal/agent-voice -g --skill create-speech-recording-controls
npx skills add yoav0gal/agent-voice -g --skill spoken-response-controls

Remove the handler with agent-voice controls uninstall.

CLI

The CLI provides small primitives that agents can combine. Run agent-voice COMMAND --help for the full options of any command.

Speak

# Positional text
agent-voice speak "The build is finished."

# Agent output through stdin
printf '%s' "$TEXT" | agent-voice speak --label build-summary

# Use a file for a long spoken response
agent-voice speak --label response < "$RESPONSE_AS_TEXT_FILE"

# Choose the output and delivery
agent-voice speak "Here is your summary." \
  --voice bf_emma --speed 1.2 --format mp3 -p

speak prints one JSON receipt with the absolute recording path, generation state, audio metadata, playback state (started or scheduled), and available delivery links. By default it returns live links while generation continues; pass --wait when the completed file is required before the command returns.

Speech text and the viewer transcript collapse whitespace, literal \n, \r, and \t escapes, and HTML space entities to spaces. Paragraph breaks are flattened. Write code and paths in their spoken form when needed; a lowercase escape in a path such as C:\new is also treated as whitespace.

Configure defaults

# Show current defaults
agent-voice config

# Set your preferred voice, speed, format, service timeout, and output directory
agent-voice config --voice bf_emma --speed 1.15 --format mp3 \
  --service-timeout 10 --output-dir ./recordings

# Restore built-in defaults
agent-voice config --reset

Voice, speed, format, and output directory can be overridden per recording with speak.

For recordings generated on another machine, configure --viewer-base-url as described below. --reset also restores localhost delivery links.

Manage the Agent Voice service

By default, speak starts the Agent Voice background service when needed. Once the model weights are loaded, the service keeps them warm between requests to avoid another cold startup. It stops after 10 idle minutes by default, and each completed speech request restarts that timer. Automatic startup reuses a running service without changing its timeout; service start uses the saved timeout or an explicit --idle-timeout value.

agent-voice service start                    # stops after 10 idle minutes
agent-voice service start --idle-timeout 30  # set this process to 30 minutes
agent-voice service stop

Use agent-voice serve for a foreground service while debugging.

Discover and diagnose

agent-voice voices
agent-voice models
agent-voice doctor
agent-voice doctor --json

voices groups each installed voice under its supported --lang tag. Select a pair with agent-voice speak --lang TAG --voice VOICE "Text".

Use --json with voices, models, config, or doctor when another tool or agent will consume the result.

Play and view recordings

agent-voice play "/absolute/path/recording.mp3"
agent-voice play "/absolute/path/recording.mp3" --after 10
agent-voice viewer start
agent-voice viewer stop

The lightweight viewer starts automatically when needed and serves only local recordings. It prefers http://127.0.0.1:8779 and selects a free port if that port is unavailable. Each managed recording keeps an editable .txt source beside it. At startup and every hour, the viewer removes temporary live-stream audio older than one hour and owned recordings older than four days and 18 hours; files without Agent Voice source, transcript, and language metadata are left alone. Opening a player or audio URL regenerates missing audio from its source with the original language and current voice and speed.

Playback commands return as soon as local playback starts, or immediately with scheduled when a delay is requested; they never wait for the recording to end.

Listen to recordings generated on another machine

Use Agent Voice 0.10.0 or newer. Run agent-voice update to upgrade an existing installation. Run Agent Voice on the remote machine and expose its recording viewer through a private HTTPS reverse proxy, such as Tailscale Serve. Configure one origin:

agent-voice config --viewer-base-url https://mini.your-tailnet.ts.net:8443
agent-voice viewer start --json

The viewer report includes its actual localhost port and its delivery_url. Proxy that port, preserving the viewer's root paths. For example, if the report says port 8779 and HTTPS port 8443 is unused:

tailscale serve --bg --https=8443 http://127.0.0.1:8779
agent-voice speak --label "Remote update" --wait "The remote recording is ready."

Use the returned delivery.browser_url as a Listen link in Codex, T3 Code, or another chat. The browser plays on the device opening the link. Remote file paths, agent-voice:// controls, and --play target the generating machine. The skills use the browser link when generation is remote.

The configured origin must use HTTPS, with no credentials, path prefix, query, or fragment. It controls returned player/audio/stream links and the exact Host accepted for GET/HEAD requests. The viewer still binds to localhost. Speech generation and local playback stay on localhost. The proxy must preserve the original Host header or add X-Forwarded-For so proxied requests cannot start playback on the host's speakers. Tailscale Serve meets this requirement. Changes to this setting apply to the running viewer without restarting it.

The setting applies to every recording on that machine, including local chats.

Serve restricts access to your tailnet; the viewer has no separate login. Anyone allowed to reach this route can read its recordings and narration text. Use a dedicated private Serve port, not a public Funnel route, and preserve existing routes. The setting does not configure Tailscale or verify remote access. Check the player link from your receiving device. If the viewer chooses a new port after a restart because its usual port is busy, update the proxy target.

Restore localhost delivery with:

agent-voice config --viewer-base-url default

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 API binds to localhost. speak uses it when available and falls back to embedded inference.

Platform limitations

Prebuilt dependencies support macOS arm64/x64, Linux x64, and Windows x64. Linux arm64 requires a C build toolchain for miniaudio. On Windows arm64, use x64 Python under emulation.