pst

Every text-to-speech option worth using sends your text to somebody else’s server and bills you per character. pst runs the whole thing on your Mac instead — no cloud, no API keys, no metering — using the Kokoro-82M model on Apple’s MLX framework. It ships as a Claude Code plugin, so asking the assistant to “read that back to me” speaks it aloud through /pst:tts.

Features

• 🔒 Fully on-device — speech is generated locally through mlx-audio; nothing leaves your machine and nothing costs money to run.
• 🗣️ 28 voices — American and British, female and male, from af_heart to bm_george, with the language inferred from the voice name.
• ✴️ Claude Code plugin — installs from its own marketplace and exposes /pst:tts, which the agent triggers on its own when you ask to hear something.
• 🔁 Always-on mode — /pst:tts always speaks a summary at the end of every turn, and verbose narrates the work as it happens. Session-scoped, cleared when the conversation ends.
• ⚡ Streamed playback — audio plays as it generates, straight to afplay, with nothing written to disk.
• ⚖️ Quality comparison — tts-compare renders the same sentence across every Kokoro quantization so you can pick the trade-off by ear.
• 🧰 One-command setup — /pst:setup installs the engine via uv and is idempotent, so it doubles as a repair command.

Get started

Add the marketplace with /plugin marketplace add joshmedeski/pst, install pst@pst, then run /pst:setup once. It needs macOS on Apple Silicon and uv. Voice modes, the CLI flags, and the roadmap toward speech-to-text and a full duplex voice loop are all on GitHub.