I'm a solo developer who lives in Claude Code and Cursor. The part that kills me isn't deciding what I want — that takes five seconds — it's the two minutes of typing that turn it into a prompt an agent can run: the goal, the numbered steps, the constraints. Plain dictation hands it back in the order I said it: fine for a message, useless as a prompt, so I'd reformat it myself. So I built ListenIn. Hold fn, ramble, let go, and the cleaned text lands at your cursor in whatever app is in front. In Claude Code, Cursor or a terminal it arrives shaped as a prompt: goal first, steps numbered, constraints in their own block. In Mail it reads like an email. Same mouth, different shape — reordered, never invented. I think in Chinese and code in English: Chinese grammar, English nouns like component, deploy, Vercel. The recognizer kept hearing "Claude" as "clock", so a dictionary and the cleanup pass repair those, and words I fix by hand join that dictionary. And if the network drops, the raw transcript still lands in the box. Losing what you just said is the one failure I refuse to ship. Recognition runs on your Mac; only text goes to a server, never audio. macOS 26 uses the new engine, older versions fall back to the classic one. Meeting recording is in there too, as part of Pro. Version 0.2.3, one person, I use it daily. Tell me what breaks.

ListenIn

Mac voice input: prompts in Claude Code, emails in Mail

Mac voice input: prompts in Claude Code, emails in Mail