Voice dictation for coding on a Mac

Nobody dictates code. What you dictate is everything around it that is written in plain language: the prompt to Cursor or Claude Code, the task you hand to Codex, the commit message, the pull-request description, the comment that explains why. That is most of the typing in vibe coding, and it is where speaking beats the keyboard.

Say the intent, type the names

Speech recognition writes words. Say “use state” and you get use state, not useState; say a file path and you get a sentence. So talk the way you would brief a colleague — what is wrong, what you want, what to leave alone — and type the identifiers, paths and flags yourself, or paste them. A two-minute spoken brief with three typed names is still far less typing than the brief alone.

In Cursor and VS Code

• Click into the chat box or the editor first, then dictate. Vocula pastes with ⌘V where the caret is, so it works anywhere ⌘V works — the agent chat, the inline prompt, a commit message field.
• Cursor and VS Code are Electron apps, which answer the accessibility system less reliably than native ones. If the text does not arrive, click into the field again and retry — the focus, not the dictation, is what went missing.
• VS Code also has Microsoft’s own VS Code Speech extension for its chat. It works inside VS Code only; a system-wide dictation app covers the terminal, the browser and everything else as well.

In Claude Code, Codex and other terminal agents

• A dictated prompt never sends itself. Vocula folds the transcript to one line before pasting, so nothing in it carries a Return. The text lands at the prompt; you read it, fix a name, and press Return yourself. The same is true at a plain shell prompt: a dictated command does not run.
• Turn off Secure Keyboard Entry. Terminal.app and iTerm2 both have it in their app menu. While it is on, macOS blocks pasted keystrokes into that window and dictation silently does nothing — how to find what is holding it.
• Careful in vim and tmux. In vim’s normal mode letters are commands. What protects you there is the terminal’s bracketed paste, not the dictation app — measured: Terminal.app does it, tmux does not. Dictate in insert mode.

What stays private, and what does not

With Vocula your voice is turned into text on your Mac, and the recording goes nowhere. The text is another matter: when you press Return, your prompt goes to Anthropic, OpenAI or Cursor exactly as if you had typed it. A local dictation app keeps your voice local. It cannot keep your prompt local.

The limits

• No custom vocabulary. Library and product names the model has not heard can come out spelled as they sound.
• No live preview. The text appears when you let go of the key, not word by word.
• Up to three minutes per press — enough for a long brief; split anything longer.
• Apple Silicon and macOS 26 only.