Over the last few releases the notch became mission control: your agents report there, you answer questions there, you approve permissions there. But starting work still meant leaving it, opening a terminal, finding the right folder, typing the launch command. And the voice side had a blind spot: a callout would happily talk over your dictation, your meeting, your voice chat.
Version 2.5 closes both gaps. New coding sessions now start from the notch itself and land back in it, ready for your first prompt. And a single microphone setting teaches every callout whose turn it is to speak.
A new session, without a terminal
Starting a coding session takes exactly two choices: pick a folder, pick an agent. Claude Code, Codex, Cursor Agent, Pi, Gemini CLI, and Antigravity are all supported, and each one launches with its own existing local settings and sign-in, so there is nothing to configure and no new account to connect. The moment the session is live, the notch announces it with Session started, ready for your input. Hover, and you can send it your first prompt right there, the same way you reply to any session.
The terminals themselves are never far. Every session you start this way lives in a sessions window with real terminal tabs: click a session on the board and it opens the right tab, or switches to it if the window is already up. From there you can stop a session, remove one you are done with, or, for Claude Code, Codex, Cursor Agent, and Pi, resume a past conversation exactly where it left off, in its original folder, with the agent's own resume machinery. The window is for managing; the notch is for working.
Callouts that wait their turn
If you dictate with tools like Wispr Flow, or spend half the day in meetings, you know the failure mode: an agent finishes, the callout starts, and now two voices are talking at once. Version 2.5 adds one switch in Playback: Wait while the microphone is active. Turn it on and automatic callouts hold while any app is using your mic, then play in their normal order once it frees up. The notch shows the waiting callout as a card, Ready, waiting on your microphone, and tapping it plays immediately, because an explicit click is you deciding it is your turn to listen.
The setting also works mid-sentence. Start talking while a callout is speaking and it pauses within about half a second, then resumes on its own when your microphone goes quiet. Your intent always outranks the automation: click and hotkey playback start immediately even while you are talking, a callout you start by hand keeps speaking until the next time you begin, and a pause you make yourself is never resumed for you.
Tuned so it never gets in your way
The switch is off by default, and if you live in voice chat all day, you can simply leave it off and keep the previous behavior. Detection reads which apps currently hold microphone input, nothing more: Agent Callout never opens your microphone, never records anything, and never asks for microphone permission.
The details are tuned from real use. macOS runs its own ambient speech services that touch the microphone on their own schedule; those never count as you talking. And if the system ever reports something unreadable, callouts play rather than wait, so a platform hiccup can never silently strand your queue.
Product decision
The principle behind 2.5
Mission control has to work in both directions. A surface where agents only report to you is a dashboard; the notch earns its place by also being where work starts, where questions get answered, and where the next prompt gets typed, without a hunt through terminal windows in between.
And audio has to earn the interruption. That has been our standard since the first callout, and 2.5 applies it to the most obvious voice in the room: yours. When you are speaking, the product waits. When you stop, it picks up exactly where it paused. Software that talks should also know when to listen.