How to Talk to Cursor by Voice (Chat and Composer)

Short answer: Cursor has no dedicated voice mode, so to talk to it by voice you run a system-wide voice-to-text app on your Mac. Put your cursor in the Chat or Composer box, hold a shortcut, speak your prompt, and the transcribed text appears in the field. An on-device tool like BlaBlaType does this in any app and keeps your audio local.

Does Cursor have a voice mode?

Not a dedicated one. Cursor is an AI-first code editor, and its Chat and Composer panels are text fields you type into. There is no official microphone button that turns your speech into a prompt. That sounds like a limitation, but it is actually freeing: because the prompt box is just a normal text field, any voice-to-text tool that types where your cursor sits can drive it. You are not waiting on Cursor to ship a feature.

This is the same trick that lets you talk to ChatGPT with voice on a Mac. The AI does not need to understand audio. A separate app handles speech recognition, converts it to text, and hands clean words to whatever window is in front. Speech recognition itself is a mature field, as the overview of the technology shows, so accuracy on modern models is high.

Talk to Cursor by voice: the setup

The whole flow takes a couple of minutes to set up and then becomes muscle memory. Here is how to dictate prompts into both Chat and Composer.

• Install a system-wide dictation app. Download BlaBlaType for macOS and grant it accessibility and microphone permissions so it can type into other apps.
• Open Cursor and focus the prompt box. Press the Chat shortcut or open Composer, then click into the input field so the cursor is blinking there.
• Hold your dictation shortcut and speak. Describe the change you want in plain language: which file, what behavior, any constraints. Release the key when you are done.
• Review and send. The cleaned text appears in the box. Read it, tweak a word if needed, and hit enter to run the prompt in Cursor.

Because the app types wherever the cursor is, the exact same steps work in Composer's multi-file view, in a terminal, in a commit message field, or in a code comment. The pattern generalizes far beyond Cursor. It is the same reason people use voice to dictate prompts to Codex CLI and to talk to Gemini CLI on a Mac.

Why voice fits AI coding so well

Prompts for AI coding tools are unusual. They are long, descriptive, and full of intent: "refactor this function to accept an options object, keep the existing default behavior, and add a JSDoc block." Typing that is slow and breaks your flow. Speaking it is natural, because you are describing what you want the way you would explain it to a colleague. Most people speak around three to four times faster than they type, and that gap is widest exactly when the text is long.

The catch is that raw speech is messy. It has "um" and "so basically" and no punctuation. That is where on-device AI cleanup earns its place: it strips filler, adds punctuation, and turns a rambling sentence into a tight instruction before it ever reaches Cursor. A cleaner prompt is a better prompt, and the model on the other end reads it more reliably.

Approach	Types into Cursor	Audio stays local	AI cleanup
On-device voice-to-text app	Yes	Yes	Yes
Cloud dictation service	Yes	No	Often
Built-in Mac dictation	Yes	Mixed	No
Copy from a transcription tool	Manual paste	Varies	No

The privacy column is not a small detail for developers. Your prompts often describe proprietary code, internal APIs, or unreleased features. With a cloud dictation service, that audio is uploaded to be transcribed. With an on-device tool, your microphone audio and transcript never leave the Mac, so only the final text you choose to send reaches Cursor. If you want the deeper comparison, we cover privacy and price across popular options, and separately look at whether Mac dictation is actually private.

Do and do not: dictating good Cursor prompts

Voice makes prompting faster, but a few habits make the difference between a prompt Cursor nails and one it misreads. Modern local models, many built on Whisper and similar architectures, handle technical vocabulary well, especially once you teach them your project's names.

Do	Do not
Add function, file, and library names to a custom dictionary so they transcribe correctly.	Assume the model knows your internal class names on the first try.
Speak the full intent in one pass, then let AI cleanup tidy the punctuation.	Stop and restart every few words trying to be perfect out loud.
Read the transcribed prompt before hitting enter in Cursor.	Send the raw text blindly and hope the AI guessed right.
Use natural language: describe the change and the constraint.	Try to dictate exact code syntax character by character.
Keep audio on-device when prompts describe private code.	Upload sensitive prompt audio to a cloud service without checking.

One more tip: a custom dictionary is the single biggest accuracy upgrade for coding work. Once BlaBlaType knows how to spell your repo names and library terms, dictated prompts read like you typed them. You can compare plans and features on the pricing page.