dsh-kokoro

If you are already running a Kokoro TTS server locally and wish the DeepSeek Harness model could actually speak instead of only printing text, dsh-kokoro is the thin function plugin that closes that gap. It ships no model weights, needs no API key, and has zero runtime dependencies.

It registers a single model-facing tool, kokoro_tts, which accepts text, voice, and speed, POSTs once to the /tts endpoint of your local Kokoro server, writes the returned WAV into a plugin-owned subdirectory under the system temp folder (directory 0700, files 0600), and hands the path back to the caller. Nineteen allowlisted voices, a 0.5 to 2.0 speed range, a 30-second abort deadline, a 10 MiB response cap, and a RIFF/WAVE header check all happen on the plugin side so a bad argument fails locally without burning a network round trip. The output path is fixed by the plugin; the caller cannot choose it, preventing text-to-speech from becoming an arbitrary file-write primitive.

It suits engineers who are already serving kokoro-tts-api on 127.0.0.1, want Harness conversations to produce audio with one tool call, and prefer a minimal attack surface with a fully offline test suite. It is not an audio toolkit: no playback, no conversion, no mixing. It answers one question, what would this text sound like, and returns a file path.