Dictate anywhere.
Stay in flow.

01 · Mac overlay

Hold Option. Talk. Keep coding.

A global Option-hold (configurable) brings up a Continuum voice overlay - waveform, partial transcript, silence countdown - without leaving the app you’re in. Release to stop, or let silence auto-stop finish the take.

There’s also an in-composer mic when you’re already in Code. Same engine, same models, same silence settings. Web and desktop composers reach the same stack through the voice bridge when a host is paired.

Partial text streams while you speak, so you can see the recognizer keeping up and stop early when it has enough.

global hotkey in-composer mic voice bridge

Code workbench Sessions

Code

composer mic

Denver · codexlocal · plan

Ready when you are.

refactor the sidebar sort so new sessions don’t jump the repo order

recording ⌥ release to send

mic active · silence 2.5ssame engine as overlay

Pick the key and the gesture in Settings → Voice. The global trigger can be left Option, right Option, or Fn / Globe, and it can fire on hold, on double-tap, or on either. Hold is push-to-talk: speak while the key is down, release to send. Double-tap is hands-free: one tap opens the take, a second closes it, and going quiet closes it for you.

The global trigger runs on a keyboard event tap, so macOS asks for Accessibility the first time you arm it. Continuum watches that trust state live and re-arms the moment you flip the switch in System Settings, with no relaunch. Leave the trigger unset and no tap gets installed at all - you still get the in-composer mic and Control-M whenever Continuum is in front.

Where the words land depends on what is frontmost. Continuum in front streams into the active composer. Another app in front, with system-wide dictation on, inserts into whatever field has focus: a commit message, a ticket, a Slack thread. System-wide off is the kill switch - the gesture then fills Continuum’s own fields only and never types into someone else’s window. A plain terminal Claude Code session has no composer to dictate into at all, which is the gap this closes.

The catalog is short on purpose. Parakeet V3 is the recommended local engine at roughly 478 MB: it runs on the Apple Neural Engine, auto-detects across 25 European languages, and is the fastest thing on the list. Parakeet V2 is English-only with the highest recall. Apple Speech is built in and downloads nothing. WhisperKit fills in the rest of the curve, from tiny near 39 MB for rough notes up through medium near 769 MB. Anything past 500 MB warns you before it starts pulling, and every card shows its size, language reach, and a coarse speed-versus-quality read.

A model downloads once and every dictating surface reuses it. After that a take never touches the network: no upload, no queue, no vendor in the path. That matters more the closer your prompts sit to production, which is the same instinct behind keeping an agent’s blast radius local. If auto-detect guesses your language wrong, pin the recognition locale in the same pane.

Availability is host-authoritative, not a hardcoded list: the client asks the host which speech models it can actually serve and renders exactly those. That keeps the picker honest across a fleet where one Mac has an xAI key and a Linux box does not. It is the same rule Continuum follows for every provider credential, and a large part of why a client beats running the CLI on its own once you have more than one machine.

The mechanics explain the one requirement. An iOS keyboard extension cannot open the microphone. So the keyboard asks the Continuum app to record, the app publishes recognition state through a shared App Group, and the keyboard reads that state and types the result into the field you were already in. That bridge is what Full Access buys, which is why the dictation key only appears once you grant it. Without Full Access the keyboard still loads, as a plain local QWERTY, and says so on its face instead of failing silently.

A finished transcript belongs to the field first. The keyboard gets a short window of first refusal to insert it; if focus moved, iOS recycled the extension, or Full Access was revoked mid-take, the text goes to the clipboard instead. It never lands nowhere - a dictation that visibly did nothing is the one outcome the delivery path exists to rule out. Dictating into Continuum’s own composer runs the same path, which is how you open a run from a phone while you are keeping several agents moving at once. An editor-bound tool like Cursor has nothing to offer the phone in your pocket.

Recording runs in the background under the system’s audio-recording indicator, so the orange dot is on whenever Continuum is listening and off the instant it stops. This is what makes voice useful for real agentic work instead of a demo: the interesting prompts arrive while you are walking to a meeting, and a thought you can only capture at a desk is usually a thought you lose.