dsh-audio-input

Typing every word into the DeepSeek Harness Web GUI is functional but slow. dsh-audio-input drops a microphone button into the message composer so you can simply speak. Partial transcripts stream live into a strip above the input box; click the button again to stop, and the final text lands in your draft, ready to send. All recognition runs on the native Apple Speech framework (SFSpeechRecognizer) on a Mac, with no cloud API key, no external service, and no audio ever leaving the machine that captured it.

Under the hood the plugin is three thin layers: the browser renders the mic button and the live-transcript strip, a cordis server-side plugin relays start and stop commands and transcripts over small JSON routes, and a roughly 130 KB native helper binary handles microphone capture via AVAudioEngine and speech recognition via SFSpeechRecognizer. Two operating modes cover different setups. In local mode the helper runs on the server Mac and uses its microphone. In remote mode the helper runs on whichever Mac you are sitting at, so both capture and recognition stay on that machine and only plain text crosses the network. The browser detects a connected helper automatically and falls back to local mode when none is present, so there is no manual toggle.

This plugin is aimed at Mac users who compose long prompts or replies in the Harness Web GUI and would rather talk than type. It is especially handy when you access a shared Harness instance from a second Mac, or when you simply prefer hands-free dictation as part of your daily workflow. No cloud subscription, no API key, just a mic button and your voice.