kokoro-cli 0.2.0

Offline-first Kokoro text-to-speech CLI and localhost service

Kokoro CLI

A small, offline text-to-speech tool for people and AI agents. It runs Kokoro-82M locally through ONNX, creates WAV, MP3, Opus, or M4A recordings, can play them on the host machine, and exposes an optional localhost HTTP API.

Kokoro CLI currently targets macOS and Linux.

Install globally

Kokoro CLI supports Python 3.11 through 3.13. Install it from PyPI with uv to keep the command in an isolated tool environment:

uv tool install kokoro-cli

pipx works too:

pipx install kokoro-cli

To upgrade an existing installation:

uv tool upgrade kokoro-cli

If uv reports that its tool directory is not on PATH, run:

uv tool update-shell

Open a new terminal, then prepare and verify Kokoro:

kokoro setup
kokoro doctor --json
kokoro speak "Hello. This is Kokoro speaking locally." --play --json

kokoro setup downloads the compact int8 model and voices once, approximately 121 MB total. Downloads are locked for concurrent processes and verified with pinned SHA-256 hashes. Speech synthesis is offline after setup.

The global kokoro command works from any directory for the current user. To install the latest unreleased GitHub version instead:

uv tool install --force "git+https://github.com/yoav0gal/kokoro-cli.git"

Optional system tools

• Install ffmpeg for MP3, Opus, M4A, and speech speeds above 2x. WAV works without it.
• Use the built-in afplay on macOS or install ffplay for local playback.

On macOS:

brew install uv ffmpeg

On Linux, install uv and ffmpeg with the methods appropriate for your distribution.

CLI

# Text argument, then play it
kokoro speak "The build is finished." --play

# Agent-friendly stdin and JSON result
printf '%s' "Here is your summary." | kokoro speak --format mp3 --json

# Select voice, speed, and output codec from the file extension
kokoro speak "A slower reading." --voice bf_emma --speed 0.85 -o recording.opus

# Go beyond Kokoro's native limit with pitch-preserving tempo adjustment
kokoro speak "A very fast reading." --speed 4.0 --play

# Discover voices
kokoro voices

# Check runtime, models, codecs, playback, and optional service health
kokoro doctor --json

# Higher-fidelity model variants with larger downloads
kokoro setup --model fp16
kokoro setup --model full

Run kokoro <command> --help for the complete option list.

Generated files use the platform user-data directory unless --output is supplied:

• macOS: ~/Library/Application Support/kokoro/recordings
• Linux: ${XDG_DATA_HOME:-~/.local/share}/kokoro/recordings

Set KOKORO_HOME, KOKORO_MODEL_DIR, or KOKORO_RECORDING_DIR to choose explicit locations.

The default int8 model is the lightweight choice; fp16 and full are available for A/B listening. Speech speed ranges from 0.5 to 4.0. Speeds through 2.0 are generated directly by Kokoro; higher speeds use FFmpeg to accelerate the result while preserving pitch.

Service selection

speak has explicit service semantics:

• --service auto is the default. It uses a healthy localhost service and falls back to embedded inference.
• --service required fails if the configured service is unavailable.
• --service off always uses embedded inference.

Set KOKORO_SERVICE_URL or pass --service-url to select another localhost port. Remote and HTTPS URLs are rejected because the service is intentionally local.

Local service for agents

Start the service:

kokoro serve

Use it from another terminal:

curl -sS http://127.0.0.1:8765/health

curl -sS http://127.0.0.1:8765/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{"input":"Your agent has finished the task.","voice":"af_heart","response_format":"mp3"}' \
  -o agent-message.mp3

The API also accepts "play": true to play the result on the host. GET /health returns the service identity, Kokoro CLI version, active model variant, and readiness. Successful speech responses include X-Kokoro-* headers for voice, sample rate, duration, generation time, and playback state.

Agent skill

The repository includes the standalone read-aloud skill. Install the CLI globally first; the skill invokes kokoro from PATH and does not depend on the repository location.

Copy the skill folder into your agent's skills directory. For Codex:

mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
cp -R skills/read-aloud "${CODEX_HOME:-$HOME/.codex}/skills/read-aloud"

Restart the agent after installation if it does not reload skills automatically.

Use from a checkout

The repository wrapper remains available for development without a global installation:

git clone https://github.com/yoav0gal/kokoro-cli.git
cd kokoro-cli
./kokoro setup
./kokoro doctor --json
./kokoro speak "Running from a checkout." --json

The wrapper creates an isolated environment through uv and keeps models and recordings inside that checkout.

Architecture

agent or human (visible text only)
    │
    ├── kokoro speak ── health check / fallback ──┐
    │                                             │
    └── POST /v1/audio/speech ────────────────────┤
                                                  ▼
                         one shared SpeechEngine
                         Kokoro ONNX + voice styles
                                    │ 24 kHz mono samples
                                    ▼
                         WAV writer / ffmpeg codecs
                                    │
                         file, HTTP bytes, or playback

The HTTP layer uses Python's standard library. Model inference is serialized behind a lock because a single local voice service values predictable resource use more than request fan-out.

The directly openable product record at docs/capabilities.html distinguishes current, verified, planned, and out-of-scope behavior.

ygent integration

Kokoro CLI exposes a stable machine-facing contract for ygent and other orchestrators:

kokoro setup --model int8
kokoro doctor --json
printf '%s' "Visible text" | kokoro speak --format mp3 --json
kokoro voices --json

The repository-owned integration manifest is integrations/ygent.json. The complete adapter contract, discovery rules, exit behavior, and JSON receipt fields are documented in docs/ygent-integration.md.

Boundaries

Kokoro CLI does not expose a remote server, cloud TTS, or speech-to-text. Other products can compose with the public CLI and localhost HTTP contracts documented above.

Model and licenses

• Kokoro-82M weights: Apache 2.0.
• kokoro-onnx runtime: MIT.
• Model assets come from the model-files-v1.0 release of thewh1teagle/kokoro-onnx.