How to Voice Type in Cursor (Agent Chat, Editor and Terminal)

Talking to an AI coding agent is faster than typing to it. Prompts get longer and more specific when you say them out loud, and a specific prompt gets better code back. Cursor supports voice input, but only in one place. This guide covers:

1. Cursor's built-in voice input, and what it can and can't do.
2. How to dictate everywhere else in Cursor: the editor, comments, commit messages and the terminal.
3. How to phrase spoken prompts so the agent gets them right the first time.

Option 1: Cursor's built-in voice input

Cursor added Voice Mode in version 2.0 (October 2025) and improved it in version 3.1 (April 2026).

How to use it:

• Click the microphone icon in the chat input box, then speak.
• In the Agents Window, you can also press and hold Ctrl + M while you talk and let go when you're done. Cursor records the whole clip, transcribes it, and shows a waveform with cancel and confirm buttons.
• You can set custom submit keywords, so saying a chosen word sends the prompt to the agent without touching the keyboard.

What it doesn't do: Cursor's voice input fills the agent prompt box only. You can't use it to dictate into the code editor, a comment, a commit message, the integrated terminal, or any other app you switch to.

For many people that's enough. If you also want voice in the rest of your workflow, add a system-wide option.

Option 2: Your operating system's dictation (free)

Both macOS and Windows have dictation built in, and it works in any text field, including Cursor's editor and terminal.

• Mac: turn it on in System Settings → Keyboard → Dictation, then press the dictation shortcut. On Apple Silicon Macs, general dictation is processed on your device.
• Windows: press Win + H to open voice typing. It needs an internet connection.

Linux has no built-in system-wide dictation, so on Linux you'll need an app (see option 3).

The trade-off: built-in dictation types exactly what you say. That means "um"s, missing punctuation and misheard technical words ("Supabase" can come out as "super base") all need fixing by hand. For short edits that's fine. For a 200-word prompt, the clean-up eats most of the time you saved.

Option 3: A system-wide AI dictation app

Apps like Blip AI sit between those two options. They work in every text field, like OS dictation, and also clean up the text before it's inserted.

Here's how it works with Blip in Cursor:

1. Install Blip from the download page and grant the microphone and accessibility permissions it asks for.
2. Click where you want the text to go. That can be the agent chat, a code comment, the commit message box, or the terminal running Claude Code or another CLI agent.
3. Hold Fn on a Mac, Ctrl + Win on Windows, or Ctrl + Super on Linux, and speak. Let go when you're finished. For longer prompts, use hands-free mode (add Space to the same keys, e.g. Fn + Space on Mac) so you don't have to hold anything down. You can change both shortcuts in Blip's settings.
4. Add your project's vocabulary to Blip's dictionary once, so it spells your framework, library and table names correctly every time.
5. Choose cloud or on-device. Blip Cloud is the default. If you'd rather your audio never leaves your machine (for example, when talking through proprietary code), download an on-device Whisper model instead. It works offline and is available on every plan, including free.

The main advantage: it's the same shortcut in Cursor, VS Code, your browser, Slack and your terminal, so you never have to think about which tool supports voice.

How to speak a prompt that works first time

Dictation makes it easy to say a lot. These habits keep what you say useful:

• Lead with the goal. "I want to add rate limiting to the login endpoint" is better than starting with background detail.
• Name the files and functions. "In auth-controller dot ts, the sign-in handler…" Specific names save the agent a search.
• State the constraints. For example: "Don't add new dependencies. Keep the existing error format."
• Say what "done" means. "It's done when a sixth attempt within a minute returns a 429 and the existing tests still pass."
• Describe the code; don't dictate it. Speaking brackets and semicolons is slow and error-prone. Say what the code should do and let the agent write it.
• Read it back before you send it. Five seconds of review catches the rare misheard word.

Here's an example spoken prompt, about 30 seconds of talking:

"Add rate limiting to the sign-in handler in auth controller. Limit to five attempts per minute per IP, and return a 429 with our standard error shape. Use the Redis client we already have in lib redis. Don't add new packages. Add a test that makes six requests and checks the last one is rejected, and make sure the existing auth tests still pass."

Typing that at an average 52 words per minute takes more than a minute. Most people speak at about 150 words per minute. We look at the research in Is voice typing faster than typing?

Which option should you use?

• You only want to talk to Cursor's agent: Cursor's built-in voice input is free and already there.
• You want voice everywhere and don't mind fixing a few words: use macOS Dictation or Win + H.
• You want voice everywhere with clean output and your own vocabulary: use a system-wide AI dictation app like Blip AI, Wispr Flow or Superwhisper. See our comparison of Wispr Flow alternatives.

FAQ

Does Cursor have voice input? Yes. Click the microphone icon in the agent chat, or hold Ctrl + M in the Agents Window. It only fills the agent prompt, not the code editor.

Can I dictate code into Cursor's editor? Yes, with your operating system's dictation or a system-wide dictation app. It works best for comments, docstrings and descriptions rather than code syntax.

Does voice typing work with Claude Code or other terminal agents? Cursor's built-in voice input doesn't reach the terminal. System-wide tools such as macOS Dictation, Win + H or Blip AI type into any focused terminal window.

Is voice input accurate with technical terms? Accuracy on technical words depends on the tool. Apps with a custom dictionary let you add library and project names once, so they come out correctly every time.