Release a seat from your licence page. Choose Show my seats, open the emailed link, and give back the seat of a Mac you no longer use. The next Mac takes it at once.
Setup can take Earshot back out of every agent, for before you uninstall. Each config file comes back as it was.
A question that arrives while you are dictating waits for you to finish. If you keep talking, the agent is told to ask again later rather than taking your words as its answer.
Model downloads pick up where they stopped, and every file is checked against its published hash before it is used.
The socket agents talk to is private to your user account. Another program on the Mac cannot open the microphone through it or answer for you.
The licence agreement inside the app now describes the monthly seat renewal the way the website already did.
0.1.1
Finds your agents when the app was opened from the Dock. An app launched that way inherits launchd's environment rather than a shell's, and launchd's PATH is four system directories: everywhere these tools install, from Homebrew to the directory Claude Code puts itself in, was missing from it. Earshot asks your login shell once instead, which is the only thing that knows about a version manager.
Registering Claude Code no longer depends on finding its command. The CLI writes one key in a file, and Earshot writes the same key the same way when the command cannot be found or refuses. Every other agent was already registered this way; this was the only one that needed an executable.
The setup window scrolls. A page with enough agents on it put the last row and its button under the footer, on a window that cannot be resized to reach them.
0.1.0
Dictation into any application, on a shortcut you choose. The words are typed into whatever already had the keyboard, so there is nothing to install per app and nothing to switch to.
Three tools any agent speaking MCP can call, over a local socket: ask a question out loud and wait for the answer, say a short summary, or put a status in the notch without saying anything.
Setup finds the agents you already have. Claude Code, Claude Desktop, Codex, Cursor, opencode, Windsurf, Zed, Cline, Gemini CLI and Hermes are registered by name; anything else that speaks MCP gets the block to paste, and a config Setup does not recognise is offered rather than ignored.
Recognition and synthesis both run on your own Mac. No account, no sign-in, and nothing about what you said leaves the machine.
Ten voices across two engines, fetched on demand rather than bundled. Every one of them was checked against the licence its training data actually carries, and three were removed on those grounds.
A different voice per project, so the agent asking you something identifies itself before you have looked at the screen.
Vocabulary corrections for the handful of words a recogniser always gets wrong for you, applied after recognition so a rule that makes things worse is visible and removable.
The last two hundred transcripts, kept on the machine and pasteable again from the menu. Free: losing what you just said is the first thing that happens to everybody, not a premium problem.
Hold the hotkey while it is speaking and it stops mid-word, immediately.
Hands-free with a wake word, and a mute that closes the device rather than ignoring it. Pro.
Updates check quietly on launch and move about forty megabytes, because the models live outside the bundle.