Voice Input for Claude Code, Cursor and AI Coding Tools

The fastest way to brief an AI coding agent is often to speak the brief. Claude Code has built-in push-to-talk dictation, but it requires a claude.ai sign-in and sends your audio to Anthropic. VS Code's built-in dictation works in chat, editors and terminals, and on desktop it can run on your computer. Cursor has a Voice Mode for its Agent, and OpenAI documents in-app dictation for the Codex desktop app but not for the Codex command-line tool. Where a tool has no built-in voice, or its voice cannot reach the prompt you are typing in, a system-wide dictation app is the fallback.

Speaking a prompt is faster than typing one, and agents reward long prompts

In a controlled study on an iPhone, English text entered by speech reached 153 words per minute, against 52 words per minute on the built-in keyboard ^{1 2}. The study used short messages in a quiet room, so it does not measure long technical prompts on a laptop. A field study of typing in everyday apps reported a different measure: in a three-week study of 30 people typing in their own apps, the mean speed was 32.1 words per minute ³. That count covers single-character keystrokes only, so it is not the same measure as the finished-text speeds above.

The two studies used different people, devices and tasks, so read the chart as a direction, not a precise ratio. The direction still matters for agents. A coding agent works from the words you give it: the files involved, the error you saw, the constraint you need kept. Speaking makes a long prompt cheap to say, so the detail is easier to keep. For a general comparison of speaking and typing rates, see our guide to speaking versus typing speed.

Claude Code has built-in push-to-talk dictation with documented limits

Claude Code's dictation is turned on with /voice in the CLI ( Claude Code docs). It has two modes. Hold mode is the default: you hold Space while you speak, and release it to stop. Tap mode starts on one tap and sends on a second tap. Your speech appears in the prompt as you talk, and you can mix typing and dictation in the same message.

The steps, for Claude Code on macOS:

1. Run /login and sign in with your claude.ai account.
2. Run /voice to turn dictation on. On a Mac, this triggers the macOS microphone prompt for your terminal if you have never allowed it.
3. If you denied access, open System Settings > Privacy & Security > Microphone, turn on your terminal app, and run /voice again ⁴.
4. Hold Space, speak, and release. Check the transcript, then press Enter. In hold mode, the transcript waits for Enter unless you set auto-submit.

If the command does nothing, see why Claude Code voice is not working.

Dictation needs a claude.ai sign-in and a local microphone

Claude Code's speech-to-text service is available only with a claude.ai sign-in ⁵. It does not run when Claude Code is set up with an Anthropic API key or a third-party cloud provider. An API key in your environment takes priority over a claude.ai sign-in, so the Claude Code docs advise removing it and restarting Claude Code before /voice can work ⁶. An unknown /voice command usually means you are not signed in with a claude.ai account ⁷.

The microphone must be on the machine you type on. Voice dictation does not work in cloud sessions or SSH sessions ⁸. The Claude Code VS Code extension supports the same dictation with the same sign-in rule, but it is not available in VS Code Remote sessions, including SSH, Dev Containers and Codespaces ⁹.

Dictation also does not use up your Claude allowance. Transcription does not consume Claude messages or tokens ¹⁰. The model's reply to the prompt still counts toward your usage.

Recording stops after 15 seconds of silence or two minutes

Claude Code stops a tap-mode recording after 15 seconds of silence ¹¹. It also stops after two minutes in total ¹². In tap mode, a transcript is sent automatically only when it is at least three words long ¹³, so a stray tap does not send one word. Press Esc or Ctrl+C to cancel a recording and discard its transcript ¹⁴.

Dictation covers a list of supported languages

Dictation uses the same language setting as Claude's replies, and it defaults to English. The page lists 20 supported dictation languages, from Czech to Ukrainian, with Hindi, Japanese, Korean and Spanish among them ¹⁵. If your language is not on the list, the command warns you and falls back to English for dictation.

Coding words are recognized, and project names are added as hints

Claude Code says its transcription is tuned for coding vocabulary in both modes ¹⁶. It also adds your current project name and git branch name as recognition hints automatically ¹⁷. These are vendor claims. The page does not publish an error rate, so check transcripts of your own terms before you trust them.

Voice dictation is not the same as Claude's voice mode. Anthropic's help center says that in Claude Code, dictation is available but voice mode is not ¹⁸. Dictation turns your speech into text for a prompt you still send. Voice mode is a two-way spoken conversation, which Claude Code does not offer.

For a broader view of Claude's own input options, see dictation in ChatGPT and Claude. A dedicated page on Claude Code dictation covers the setup in more detail.

Cursor, VS Code and Codex each handle voice differently

The four tools document their voice input at different levels of detail, so the table shows what each vendor has published rather than what you might guess.

VS Code's dictation reaches chat, editors and terminals

Microsoft describes VS Code's built-in dictation as converting your speech to text in chat, the Agents window, editors and terminals ¹⁹. That is the broadest built-in coverage in this comparison, and terminals are the hard case, which is covered below. To start Voice Mode, which reads the agent's replies aloud and listens for follow-up requests, select the Voice Mode button in the chat input. The VS Code voice page gives Shift+Command+Space on macOS and Ctrl+Shift+Space on Windows and Linux as the shortcut. The VS Code Speech extension adds a "Hey Code" keyword activation feature for hands-free starts.

Microsoft labels built-in dictation as experimental, and says it might change or be removed ²⁰. Intel-based Mac computers are not on the list of supported desktop platforms, so those users install the VS Code Speech extension instead ²¹. For the setup steps, see voice input in VS Code.

Cursor's Voice Mode is documented in its changelog

Cursor's changelog for version 2.0 lists Voice Mode as a way to control the Agent with your voice using built-in speech-to-text conversion ²². The same entry says you can define custom submit keywords in settings that start the Agent running ²³.

I could not open a Cursor documentation page for voice. The changelog does not say where Cursor's speech-to-text runs, and it does not list keyboard shortcuts. Third-party guides give shortcuts, but I could not confirm them against Cursor, so this guide does not list them. Check Cursor's settings for the current shortcut.

Codex's official docs cover the desktop app, not the command line

OpenAI's configuration reference includes a switch for in-app dictation in the desktop app. The setting is features.in_app_dictation, and it disables in-app dictation in the desktop app ²⁴. That is the only voice setting I found in the reference.

Third-party blogs describe a spacebar push-to-talk in the Codex command-line tool, and they disagree about whether it still ships. The official reference does not document it, so this guide does not rely on those reports. If you use the command-line tool, check its release notes for your version, or use a system-wide dictation app.

Built-in voice input sends audio to a cloud service

Where the audio goes is the first question to answer for any built-in voice input. For Claude Code, Anthropic's documentation says audio is not processed on your computer. Voice dictation streams your recorded audio to Anthropic's servers for transcription ²⁵. The same page confirms that the audio is not processed locally.

VS Code is split by platform. On the desktop, dictation uses an on-device speech recognition model by default, so you can dictate without sending audio to an online service ²⁶. The model downloads on first use, and after that dictation works offline ²⁷. In VS Code for the Web, the on-device model is not supported, so dictation streams microphone audio to the Microsoft AI voice service ²⁸.

Each row can change when the vendor updates the product, so check the vendor's page before you rely on it for confidential code. The question of whether voice data is safe in general is covered in whether voice to text is safe, and the on-device options are in on-device speech recognition.

Terminals are the hardest place for outside dictation to type into

A terminal is a hard target for dictation because it does not behave like a normal text box. Claude Code's built-in dictation has to handle the keys itself. Its hold mode works by watching for key-repeat events from your terminal, so the feature needs the terminal to send them ³¹. Some keys are not delivered to terminal applications at all, so Claude Code cannot bind its dictation key to them ³².

Outside dictation has its own limit. Windows voice typing works only when the cursor is in a text box ³⁰. Whether a given terminal exposes its prompt as a text box to the operating system depends on the terminal app and the tool running in it. I could not test every terminal, so the test to run is simple: put the cursor at the prompt, dictate one short sentence, and check that the words appear before you rely on it.

VS Code is the exception the vendor documents. Its built-in dictation converts speech to text in terminals as well as in chat and editors ¹⁹. If your agent runs in VS Code's terminal, the built-in dictation may cover it. A system-wide tool may still be needed when you run the agent in a separate terminal app. The mechanism is covered in how dictation types into an app, and the options are compared in dictate into any app.

Technical vocabulary is where dictation still makes mistakes

Coding terms are where speech recognition is most likely to fail. A 2026 arXiv preprint on spoken queries about code found that these queries combine non-standard English, domain-specific vocabulary and custom identifiers, and it names "custom identifiers such as variable and function names" as a central difficulty ³³. The study covered four Indic languages and English ³³. It used code understanding benchmarks, so it does not measure dictating prompts in your own editor.

Claude Code's dictation is tuned for coding words, and it adds your project and branch names as hints ^{16 17}. Those are vendor statements without published accuracy numbers, so do not assume a name will be right. Read the transcript before you press Enter, and correct identifiers by hand. Accuracy is usually reported as word error rate, and the speech recognition guide explains how it is calculated.

If you dictate the same names every day, a system-wide dictation app with a custom vocabulary list may help. A custom vocabulary is a list of terms the recognizer is told to expect. Check whether your app supports one before you rely on it.

For code syntax, not prose, another route is voice coding. Tools such as Cursorless and Talon Voice let you edit code by speaking structure, which is a different job from dictating a prompt.

A system-wide dictation tool fills the gaps the built-in buttons leave

Use the built-in button where it is documented to work and where its processing matches your needs. Claude Code's /voice is the documented option for that tool, with the sign-in and cloud limits above. VS Code's dictation is the widest built-in coverage. Cursor's Voice Mode is documented only at the changelog level. Codex's documented desktop dictation does not cover the command-line tool.

A system-wide dictation tool, sometimes called system-wide dictation, types into whatever text field has focus. That is the fallback for an editor, browser or terminal that the built-in button does not reach. Its limits differ: it needs a focused field that accepts typed text, and its processing depends on the product. Check the processing setting before you use it for confidential code.

Yap, which we make, is a dictation app for Mac and Windows that runs on your device. It is one option in this list, not the only one, and the built-in tools above may be enough for your work. Download the app if you want a system-wide option, or read how to choose a dictation app to compare the choices.

For a broader map of dictation across apps, the overview is dictation in your apps. Other guides in this set cover Google Docs voice typing, dictation in Word and speech to text on Linux, and the Mac dictation guide covers the built-in option on Apple computers.