speak

Make OpenAI Codex CLI speak the last assistant response on turn completion (macOS), with a SwiftBar menu bar play/stop toggle.

What it installs

• ~/.codex/say.sh: notifier script invoked by Codex (notify = ["~/.codex/say.sh"])
• SwiftBar plugin: ~/Documents/SwiftBar/codex-say.5s.sh (configurable)
• Runtime state (no secrets): ~/Library/Caches/codex-say/

Install

./install.sh

Options:

# Choose a different SwiftBar plugins folder
SWIFTBAR_PLUGIN_DIR="$HOME/Documents/SwiftBar" ./install.sh

# Skip installing SwiftBar (installer prints instructions instead)
INSTALL_SWIFTBAR=0 ./install.sh

# Skip installing uv (required for Pocket TTS via `uvx`)
INSTALL_UV=0 ./install.sh

# (Optional) warm Pocket TTS deps into uv cache
WARM_POCKET_TTS=1 ./install.sh

SwiftBar Install

If you don’t have SwiftBar yet:

brew install --cask swiftbar

Or download from:

• https://github.com/swiftbar/SwiftBar/releases/latest

Usage

SwiftBar usage

SwiftBar basics:

• Left click the menu bar title to toggle Play/Stop (or Enable when Off).
• Right click (or click the title then open menu) to open the dropdown options.

Title hints:

• Fast = Apple say engine (low latency).
• Slow = Pocket TTS engine (higher latency; CPU-dependent).
• q5 means 5 completions were queued while speech was already playing (it will speak the latest after the current one ends).

Terminal controls:

~/.codex/say.sh toggle        # enable/disable auto-speak
~/.codex/say.sh replay        # replay last spoken summary
~/.codex/say.sh replay-full   # replay full last response
~/.codex/say.sh stop          # stop speech

# Optional engines
~/.codex/say.sh engine apple  # use macOS `say` (default)
~/.codex/say.sh engine pocket # use Pocket TTS if installed (falls back to Apple if missing)
~/.codex/say.sh pocket-voice alba

SwiftBar:

• Menu bar title shows Off / Ready / Speaking
• Click the title to Play/Stop

Notes

• Speech reads an intelligent summary by default: skips code blocks, truncates very long IDs/identifiers, and stops before a Test/Tests/Testing section.
• Volume is speech-only by default (system volume unchanged): audio is generated to a single cached file and overwritten each time (no disk growth).
• Parallel turn completions are handled: if Codex finishes again while already speaking, it queues “latest” and will replay after the current speech ends.
• No credentials are stored; only the last response text and playback state.

Pocket TTS (optional)

This project can use Kyutai Pocket TTS as an alternative engine.

Install uv, then:

uvx pocket-tts generate --help
uvx pocket-tts serve

Then in SwiftBar (or terminal) select the “Slow” engine (Pocket TTS).

Pocket TTS caveats:

• This project controls Pocket volume via the audio player (afplay -v), not system volume.
• Pocket can feel “slow” because it has to generate audio before playback (CPU-dependent; expect a few to several seconds of lag on some machines).
• Pocket does not expose a speech-rate option in the CLI; this project can only speed up/down playback rate (via afplay -r), which may affect pitch/quality. For the snappiest experience, prefer Apple say (“Fast”).

License

Apache-2.0 (see LICENSE).