MicDrop

The idea

Coding agents already run for minutes at a time. You should be able to walk away, say what you want, and hear back what happened, without teaching the agent anything new.

One voice turn

you speak → Whisper transcribes on your CPU
          → pasted into the agent's chat box + Enter
          → the agent works and replies as usual
          → its after-reply hook hands MicDrop the reply
          → the first sentences are read aloud → the mic opens again

Key decisions

• Invisible to the agent. MicDrop adds one quiet hook per agent (Stop or afterAgentResponse). The hook never prints into the chat and never slows the agent down, even when MicDrop isn’t running.
• Speak the gist, not the diff. It reads only the first paragraph, with code, paths and markdown removed, and stops at a sentence boundary.
• Three ways to listen: turn by turn (the default, so room noise isn’t sent mid-task), hands-free, and push-to-talk.
• Voice commands for MicDrop itself (“stop listening”, “repeat that”, “say that slower”). They count only as a whole sentence, so they never collide with real requests.

Packaging

It’s a Rust workspace with a core, a CLI and a tray app. You can install it with npx mic-drop install, a one-line script, or an MSI/DMG/AppImage desktop build.