Agent-mode Cursor is a prompt-writing job. You are not typing code any more; you are typing paragraphs describing what the code should do, then more paragraphs about why the first attempt was wrong. Speech runs around 150 words per minute. Typing, for most people, lands near 40.
Vokk is a push-to-talk key that works everywhere on your Mac. Hold it, talk, release — the words are typed into whatever has focus. Cursor's chat box. The terminal pane. A comment in the editor. The PR description in your browser afterwards.
Cursor ships voice input for its chat, and it's genuinely convenient: it's already there, it's in the box you're typing into, nothing to install. If chat is the only place you dictate, use it.
The difference is scope. Built-in voice is a feature of one text box in one app. Vokk is one habit that works in every text box you own — Cursor's chat and its terminal and its editor, plus Slack, GitHub, your notes app, the commit message. You learn the key once instead of learning each tool's dictation button and living without one wherever it doesn't exist.
| where you can dictate | built-in chat voice | vokk |
|---|---|---|
| cursor chat / agent prompt | yes | yes |
| cursor's terminal pane | — | yes |
| the editor itself | — | yes |
| browser, slack, anything else | — | yes |
Because Vokk types into the focused window locally, it works in any terminal window regardless of what that terminal is connected to.
Dictation into a frontier model has an unusual property: the reader is very good at intent. If "useEffect" lands as "use effect", or a variable name comes out slightly mangled, the model still knows what you meant and asks if it doesn't. You are not writing a legal filing. You are handing over context.
So the last percent of transcription accuracy matters less here than two things that don't degrade gracefully: how fast the words appear, and where the audio goes.
Vokk transcribes on-device by default — whisper.cpp on your Mac's GPU via Metal, well under a second. No account, no sign-up, no telemetry. The airplane test: turn Wi-Fi off, hold the key, it types anyway. Cloud transcription via Groq exists as an option behind a license, but local is the default and local is enough.