Voice coding

Claude Code voice: what's true about talking to your coding agent

Updated November 17, 20268 min read

Claude Code voice dictation lets you speak your request out loud and have it typed into Claude Code for you. You turn it on with the /voice command, hold the Space bar while you talk, and let go to stop. Plenty of people never try it because of a few old ideas about voice coding. This page clears those up, then shows every way to set it up and how to speak prompts that get good results.

Voice coding today means describing, not dictating code

Years ago, voice coding meant reading code aloud symbol by symbol: "open bracket, const, space, user". It was slow and tiring, and mostly used by people who could not type because of pain or injury.

AI coding agents changed that. With Claude Code you do not say the code. You say what you want, in normal sentences, and the agent writes the code. That is a much better fit for speech. When you talk, you tend to explain more: what is broken, what you expected, what the page should feel like. That extra detail is exactly what an agent needs to get things right first time.

Speech to text coding is simply that loop: you talk, the words land in the agent's prompt box, you check them, and you send. Every option on this page does the same basic thing. They differ in where they work, what they need and where your audio goes.

Six myths about coding by voice

Myth: voice coding is only for people who can't type

Fact: it is a big help if typing hurts, and that matters a lot. It also suits people who type fine. Speaking a long request is less tiring than typing it, and you can do it while looking at the app you are fixing. A good habit is to talk through the long "what and why" part and type the short exact parts.

Myth: you have to say every bracket and semicolon

Fact: with an agent you describe the result. "When the email box is empty, show a red message under it" is a complete instruction. Claude works out the code. You only spell things out for exact values like a price or a web address.

Myth: dictation garbles technical words

Fact: general dictation does sometimes trip on jargon, but Claude Code's own dictation is tuned for coding words. Anthropic's docs say terms like regex, OAuth, JSON and localhost are recognised, and your project name and git branch name are added as hints automatically. Other tools let you add a custom dictionary for your own product names.

Myth: talking to Claude Code eats into your usage limits

Fact: per Anthropic's voice dictation docs, turning speech into text does not use Claude messages or tokens and does not count toward the limits in /usage. Only the prompt you send counts, the same as if you typed it.

Myth: you must get the whole prompt right in one breath

Fact: you can mix talking and typing in the same message. In Claude Code, the words appear at your cursor, so you can speak a sentence, type a file name, then hold Space again to add more. Press Esc to throw away a recording you did not like. Nothing is sent until you press Enter, unless you turn auto-send on.

Myth: your voice always goes to the cloud

Fact: it depends on the tool. Claude Code's /voice does send audio to Anthropic's servers to be transcribed, and Windows voice typing needs an internet connection. Some tools can turn speech into text on your own computer instead, which suits private projects and works without a connection.

Option 1: the built-in Claude Code /voice command

As of November 2026, voice mode in Claude Code works like this:

/voice        # turn dictation on or off (hold mode by default)
/voice hold   # hold Space to record, let go to stop
/voice tap    # tap Space to start, tap again to stop and send
/voice off    # turn it off
  • Hold mode: hold Space. After a short warm-up the footer says listening…. Your words show up as you speak. Let go, check the text, press Enter.
  • Tap mode: with an empty prompt box, tap Space once to start and again to stop. It sends by itself if you said at least three words. Recording also stops after 15 seconds of silence or two minutes in total.
  • Needs a Claude.ai login. It does not work with an API key or a third-party provider. If you see "Voice mode requires a Claude.ai account", run /login.
  • Needs a microphone on the same computer. It does not work over SSH or in cloud sessions. In WSL on Windows it needs WSLg.
  • 20 languages are supported, set through /config. English is the default.
  • Rebind the key in ~/.claude/keybindings.json if Space does not suit you, for example to meta+k, which starts recording with no warm-up.

The Claude Code VS Code extension has dictation too, with the same login rule. If Windows blocks the microphone, open Settings > Privacy & security > Microphone and turn on access for desktop apps. On a Mac, use System Settings > Privacy & Security > Microphone and switch on your terminal app.

Option 2: dictation already built into Windows and macOS

Your computer can already turn speech into text in almost any box, including a terminal running Claude Code. This is handy if you use an API key, which rules out /voice, or if you want to talk to other agents too.

Windows voice typingMac Dictation
StartPress Windows key + HPress the Microphone key, your chosen shortcut (such as pressing Fn twice), or Edit > Start Dictation
StopSay "stop listening" or click the micPress Esc, the Microphone key or the shortcut again
Turn onWorks out of the boxSystem Settings > Keyboard > Dictation
PunctuationOptional automatic punctuation in its settingsAdds commas, full stops and question marks by itself in supported languages
InternetNeeded (online speech recognition)Check Keyboard settings to see if it runs on your Mac

Both are free. Both also understand "new line" when you want a line break. The weak spot is jargon: library names and file names often come out wrong, so read the text before you send it.

To use either one with Claude Code, click into the terminal so the cursor sits in Claude's prompt box, start dictation, and talk. The words are typed in as if you used the keyboard. Be careful with a spoken "new line" here: in a terminal, a dictated line break can act like pressing Enter and sends the prompt early. If that happens, finish your sentence, then add line breaks by hand with Ctrl+J.

Option 3: dictation apps and VibeCodingDeck

Standalone dictation apps type wherever your cursor is, which makes them handy for dictation for coding in any tool. Wispr Flow is one for Mac, Windows, iPhone and Android. As of November 2026 it has a free plan with a weekly word limit, and a Pro plan at $15/month ($12/month billed yearly). Our Wispr Flow alternatives page compares it with other choices.

VibeCodingDeck has dictation built in, made for driving coding agents. Press a hotkey, talk, and the words are typed into whichever agent you are using: Claude Code, Codex, Gemini CLI or one of the others. It can turn speech into text on your own computer using whisper.cpp, sped up by your graphics card on Windows, or in the cloud if you prefer. It works system-wide too, so it types into any app, and it has a custom dictionary for words it keeps mishearing.

It pairs well with Point & Fix: click something in your app's live preview, then say what should change. VibeCodingDeck runs on Windows 10/11 and Apple silicon Macs; see features.

When /voice will not listen: quick fixes

Most voice problems come from the microphone, the login or the language setting. These fixes come from Anthropic's troubleshooting list:

  • "Unknown command: /voice": you are not signed in with a Claude.ai account, or an API key is set and takes priority. Run /login, or remove the ANTHROPIC_API_KEY setting and restart Claude Code.
  • "Microphone access is denied": give your terminal app microphone permission in your system settings, then run /voice again.
  • Holding Space just types spaces: dictation is probably off. Run /voice hold. If one or two spaces appear and then nothing, your computer may have key-repeat turned off, so switch to /voice tap instead.
  • "No audio detected from microphone": the wrong microphone is set as the default. On Windows check Settings > System > Sound > Input. On a Mac check System Settings > Sound > Input.
  • Text comes out in the wrong language or garbled: set your language in /config. Dictation falls back to English if your language is not on the supported list.
  • Linux asks for SoX: install it with the command Claude Code shows, for example sudo apt-get install sox. In WSL, use sudo apt install sox libsox-fmt-pulse.
  • "Voice mode is disabled by your organization's policy": an admin has turned it off for your company account. Ask them, or use your computer's own dictation instead.

Tips for dictating prompts that Claude gets right

  • Say the outcome first, then the details: what you want to see when it works.
  • Name the place: "on the checkout page", "in the header", "in the sign-up form".
  • Say what must not change, so the agent leaves the rest alone.
  • Type exact values (prices, web addresses, variable names) instead of saying them.
  • Read the text before you press Enter, especially numbers and names.
  • Use plan mode for big spoken requests so you can check Claude understood before it edits files.
  • Keep each recording to one job. Two short prompts beat one long ramble.
  • Add words it keeps mishearing to a custom dictionary if your tool has one.

Here are spoken prompts that work well, word for word as you might say them:

"On the pricing page, make the yearly plan the one that's selected first.
Don't change the prices or the button colours."

"The sign-up form lets me submit with an empty email. Stop that, and show
a short red message under the email box. Then test it with an empty email."

"Plan first, don't edit yet. I want a dark mode switch in the settings page
that remembers my choice. Tell me which files you'd change."

Notice none of them contains code. Each says where, what and what to leave alone. For more prompt patterns, see vibe coding prompts and our guide to plan mode. New to Claude Code itself? Start with how to use Claude Code.

Sources

Run up to 12 AI coding agents in one window.

Claude Code, Codex, Gemini, Grok and more, side by side. Windows and Mac. From $29 a month.

Get started

Keep reading