pst

pst

Local-first speech tooling optimized for macOS (text-to-speech and speech-to-text).

A local-first, on-device voice toolkit for macOS (Apple Silicon), packaged as a Claude Code plugin.

Today pst gives you text-to-speech that runs entirely on your machine — no cloud, no API keys, no per-character billing — powered by mlx-audio and the Kokoro-82M model running on Apple's MLX framework.

Installed as a Claude Code plugin, it exposes the tts skill as /pst:tts (and the agent triggers it automatically when you ask it to "read that back" or "say that out loud"). The pst: namespace is the umbrella for the voice skills on the roadmap — /pst:stt, /pst:converse, and more.

Install as a Claude Code plugin

pst is distributed as a plugin through its own marketplace (this repo). In Claude Code:

/plugin marketplace add joshmedeski/pst
/plugin install pst@pst

That's it — the tts skill is now available as /pst:tts, and the bin/ scripts (tts, tts-compare) are on the Bash PATH while the plugin is enabled.

Prerequisite: the skill shells out to the local tts binary, which needs the mlx-audio engine. Run the one-time setup first:

/pst:setup

This installs the on-device tooling via uv (see engine setup for what it does and the manual equivalent). It requires uv — the command tells you how to install it if it's missing.

Local development

To iterate on the plugin from a checkout without publishing:

git clone https://github.com/joshmedeski/pst ~/c/pst

Then point Claude Code at the local marketplace and reload after edits:

/plugin marketplace add ~/c/pst
/plugin install pst@pst
/reload-plugins        # pick up edits mid-session

Main feature: text-to-speech with mlx-audio

The bin/tts script speaks any text aloud using the Kokoro TTS model. Because everything runs locally through MLX, it's fast, private, and free to run as often as you like.

tts "hello world"
tts --voice bm_george "good evening"
tts --voices          # list available voices, grouped by language
tts --help

• 28 voices across American/British and female/male, e.g. af_heart (default), am_michael, bf_emma, bm_george.
• Streams audio live as it generates — nothing is written to disk.
• The correct lang_code is inferred from the voice prefix (a* → American, b* → British).

Voice modes: always-on and verbose

/pst:tts takes optional mode arguments so the assistant can keep talking throughout a session instead of only when asked:

Always-mode is session-scoped: it applies to the current conversation only and is cleared automatically when the session ends. State lives in ${XDG_CONFIG_HOME:-~/.config}/pst/voice/. A UserPromptSubmit hook keeps the mode active each turn (it never speaks — the assistant authors all audio), so existing installs must run /reload-plugins once to pick up the hooks.

Comparing model quality

bin/tts-compare generates the same sentence across the Kokoro quantization variants (bf16, 8bit, 6bit, 4bit) so you can pick the quality/size trade-off that sounds best to you. Each sample self-labels its variant so you can identify it by ear.

tts-compare                     # generate (if missing) and play all four
tts-compare "custom body text"  # override the sentence; regenerates
tts-compare -f                  # force regeneration even if cached

Setting up the TTS engine (mlx-audio)

The recommended path is the pst-setup script — run it in Claude Code as /pst:setup, or directly:

pst-setup

It's idempotent (safe to re-run, doubles as a repair command) and does exactly what the manual steps below do. It requires uv and tells you how to install it if it's missing.

uv tool install --force git+https://github.com/Blaizzy/mlx-audio.git \
  --prerelease=allow --with "misaki[en]" --with num2words \
  --with espeakng-loader --with "numpy>=1.26,<2" \
  --with "spacy>=3.7,<4" --with pip