Source: https://fnflow.app/vibe-coding-voice-input

# Vibe coding with voice input
The term has had a microphone attached from the start. In the February 2025 post that named vibe coding, Andrej Karpathy wrote “I just talk to Composer with SuperWhisper” and said he barely touched the keyboard. Since then, most AI coding tools have added voice of their own. Here's what's built in, when a system-wide app fits better, and how to speak a prompt an agent can use.

## What's built in

| Tool | Voice input | Where speech is recognized |
|---|---|---|
| Claude Code | /voice, hold Space | Anthropic's servers |
| Copilot CLI | /voice, hold Space | Your machine |
| VS Code | Chat microphone, ⌘I | Your machine on Apple silicon |
| Gemini CLI | Experimental /voice, hold Space | Google Cloud, or Whisper if you set it |
| Antigravity | Microphone, Ctrl+M; F5 in the CLI | Gemini Audio models |
| Codex CLI | None in the docs we read on 18 October | – |

From each tool's docs. See our pages on [Claude Code](claude-code-voice-mode), [Copilot](github-copilot-voice-input), [Gemini CLI](gemini-cli-voice-input), [Antigravity](antigravity-voice-input) and [Codex](dictation-codex-cli). Checked 6 November 2026 unless noted.

## Built-in or system-wide

Built-in voice is the quickest way to start if you stay in one tool. A system-wide app makes more sense if you switch tools during the day, because you learn one key instead of six. It also covers the rest of the work around the agent: replying in Slack, writing an issue or a commit message. [Voice input for Cursor](dictation-for-cursor) shows how this works in practice.

## How to talk to an agent

- **Start with the outcome.** “The signup form should show an error when the email is taken.” Then give the details.
- **Say the constraints out loud.** What not to touch, which library to use, how you'll check it. Voice makes this easy, so use it.
- **Ramble, then read.** Since the text lands in the input without being sent, you can say it badly and fix a word before sending.
- **Name files plainly.** Then correct the spelling, or add project names to your vocabulary so they come out right the first time.
- **Hold, don't toggle.** With push-to-talk the recording ends when your finger lifts, so a conversation in the room doesn't end up in your prompt.

[Dictating prompts for AI coding tools](dictate-ai-prompts) has more, with examples.

## A caution about the “vibe” part

Speaking faster doesn't make accepting changes unread any safer. Karpathy described his approach as fine for throwaway weekend projects. Voice helps you say more of what you mean, but on real code you still need to review the diff.

## Where FnFlow fits

FnFlow is a free Mac app built for this: hold fn, say the prompt, release, and it's pasted into whatever is in front. Whisper runs on your Mac's GPU, so about 1.2 seconds after you let go of a 7.6-second sentence (M4 Pro) the text is there and nothing has left the machine. It keeps every take with its audio, in case a long prompt doesn't land.

## Questions

**Why use voice for vibe coding?**
Because the input is mostly plain language, and it's easier to give full context out loud than to type three paragraphs. You'll say the constraints you'd skip when typing.

**Is voice input good for writing code itself?**
Not really. It's for describing what you want. Dictating syntax is slow, so let the agent write it. Voice coding on Mac covers what works and what doesn't.

**Does my voice go to the cloud?**
It depends on the tool. Claude Code's /voice streams audio to Anthropic. Copilot CLI and VS Code's desktop dictation recognize speech on your machine. Gemini CLI goes to Google Cloud by default, with a local Whisper option.

**What about the office?**
Talking aloud isn't always possible, but a headset helps, and you can keep typing for small edits. Many people save voice for long prompts.
