AI Workflow

Voice Coding With AI: Talk to Claude Code, Cursor and Copilot Chat

Much of a developer's day is now spent writing paragraphs to a coding agent. Those paragraphs are prose, and prose is faster to say than to type. With AirTypes, Plain mode turns your speech into a prompt for free, offline, on your own CPU; Polished tidies a rambling brief at 120 minutes of speech for every dollar. Here is how to dictate your prompts, what to keep on the keyboard, and how to set it up in a few minutes.

Voice coding with AI means speaking your instructions to a coding agent instead of typing them. You do not dictate the code. You describe the change, the bug or the review you want, and the agent writes the code. Here is a real one, said out loud and typed by Polished:

You say

"refactor the payment service so it uh retries failed webhooks three times with backoff and like logs each attempt"

→
Polished types

"Refactor the payment service so it retries failed webhooks three times with backoff and logs each attempt."

With AirTypes, you click into the Claude Code prompt in your terminal, the Cursor chat panel or GitHub Copilot Chat in VS Code, hold a hotkey, speak and release. The text is typed at your cursor. You read it and press Enter. It works in all of those tools for one reason: AirTypes types at the cursor in any desktop app, so there is no plugin or integration to install. If you can type a prompt somewhere, you can speak one there.

Claude Code also has its own /voice command, which suits you if you sign in with Claude.ai and only dictate there. Claude Code voice dictation compares the two.

Why voice fits prompting

Coding agents do better work with context. A useful prompt says what is broken, what you already tried, which part of the codebase is involved, what must not change, and how you will judge the result. That is five or six sentences, and when you type them, you cut corners. The prompt becomes "fix the login bug" and the agent spends its first turn guessing.

Speaking removes most of that friction. You explain the problem the way you would explain it to a colleague, reasoning and constraints included, because one more sentence costs a few seconds. You get a longer, more specific prompt for less effort, and your hands can stay on the diff or the stack trace while you talk.

If you already dictate to chat assistants, the same habits carry over. Our guide to voice prompting ChatGPT, Claude and Gemini covers the general technique, and voice to AI prompt covers the tooling; this post focuses on coding agents.

Where voice does not fit: code syntax

Dictating code character by character is awkward with any speech-to-text tool, AirTypes included. Brackets, semicolons, operators and identifiers like getUserById are slow to say and easy to mishear. Whisper is trained to produce natural sentences, so it is likely to write "get user by ID" where you wanted a camelCase name.

The practical split:

  • Say: the goal, the symptoms, the constraints, the acceptance criteria.
  • Type: exact identifiers, file paths, error codes and snippets the agent must match character for character.
  • Let the agent find it: "the click handler in the login form component" is usually enough for an agent that can search your repository.

Claude Code and Cursor both let you reference a file with an @ mention in the prompt box. Dictate the sentence, then add the reference by keyboard.

Setup: dictate to Claude Code, Cursor and Copilot Chat

Nothing changes in your coding tools. The setup lives in AirTypes.

  1. Install AirTypes

    Download AirTypes for macOS or Linux and create a free account at first launch. No card. On macOS, allow Microphone and Accessibility when asked.

  2. Pick a model tier

    Plain has two tiers, Swift and Crisp. Pick Crisp, the one we recommend. Developer prompts are full of library names, so when Crisp mishears them, switch to Polished for that prompt.

  3. Keep the default hotkey or change it

    The default is Ctrl + Shift + Space. If your editor or terminal already uses it, pick another combination in Settings.

  4. Choose Hold or Toggle

    Hold (the default) records while you hold the hotkey, good for short prompts. Toggle starts on one press and stops on the next, easier for a long brief.

  5. Put the cursor in the prompt box

    Claude Code: the prompt line in your terminal. Cursor: the chat or agent panel input. Copilot Chat: the chat input in VS Code. Then hold, speak, release, read and press Enter.

One setting matters in the terminal. The default text output, Direct Paste, simulates Ctrl + V (Cmd + V on macOS), and many Linux terminals paste with Ctrl + Shift + V instead. If you dictate into a terminal, switch Settings > Controls > Text Output to Type at cursor, which types each character and works in every app. Either way, AirTypes types the text and stops; nothing is sent until you press Enter.

Which Plain tier for technical prompts

TierWhisper modelDownloadFor coding prompts
Swifttiny (English)~78 MBFast, but misses library names often
Crispbase (English)~148 MBThe tier we recommend; fine for short, plain-English prompts

Both Plain tiers run locally on a 4 GB machine with no GPU and handle English only. When you need more accuracy than Crisp on framework names and acronyms, use Polished: a large speech model on the server transcribes the recording, so it needs a connection and uses prepaid credit after the free first minute. See Whisper model sizes explained, or the Linux setup guide and macOS setup guide.

Talk to your coding agent instead of typing to it

Plain is free and transcribes offline, typing at your cursor in the terminal, your editor or the browser. macOS and Linux.

Download AirTypes

Prompt patterns that work by voice

Three spoken prompts that give an agent enough to act on. They are written the way people talk; in Plain, AirTypes types them as you say them, with a leading "um" dropped, a capital first letter and a final period. In Polished, the filler and false starts are removed too.

The bug report

Say what happens, where you suspect the problem is, and what must stay untouched.

Say this

"The login form on mobile Safari, when you tap submit nothing happens, no error in the console. I think it's the click handler in the login form component. Can you fix it without changing the desktop layout?"

The refactor request

Constraints are what people skip when typing, and they keep an agent from touching half the codebase.

Say this

"The billing service is doing way too much. Split the Stripe webhook handling into its own class, keep the public methods the same so nothing else breaks, and add tests for the new class."

The review request

Asking for a list before any edits keeps you in control of the change.

Say this

"Before I open the pull request, review the changes on this branch. Look for anything that breaks existing behavior or missing error handling. Just list the problems, don't fix them yet."

Two habits help. Read before you press Enter: a misheard library name costs you a wasted agent turn, and a two-second read catches it. And answer the agent's follow-up questions by voice too; short replies like "yes, keep the old endpoint for now" are where dictation saves the most switching. New to dictation? The first week of a voice dictation habit shows how to make it stick.

Where Polished helps, and where Plain is the right call

AirTypes has two modes, switched on the Home screen. Both type at your cursor. They differ in where your speech is transcribed and how much cleanup happens.

Plain: proprietary code stays local

Plain runs Whisper offline on your own CPU and uploads 0 bytes of audio. Describing a bug out loud often means naming internal services, customers or unreleased features; in Plain, that recording never leaves your machine. Plain is free, works without a connection, and its light cleanup (leading "um" removed, first letter capitalized, period added) keeps your words exactly as said. That literalness is a feature for prompts: what you said is what the agent gets.

Use Plain for: anything under NDA or about proprietary code, short prompts you say cleanly, follow-up answers to the agent, and working offline.

Polished: long, thinking-out-loud briefs

Polished sends the recording to AirTypes to transcribe and clean it up. A large speech model transcribes it, then an AI removes filler words and false starts and fixes grammar and punctuation, keeping your meaning, numbers and names. It helps most when you think out loud:

You say

"so um I think we should uh probably move the deploy to friday because like the migration isn't, the migration isn't tested yet and uh QA needs two more days"

→
Polished types

"I think we should move the deploy to Friday because the migration isn't tested yet and QA needs two more days."

You say

"um can you explain why my react component re-renders every time the parent updates, I'm passing an object as a prop and uh I already tried memo"

→
Polished types

"Why does my React component re-render every time the parent updates? I'm passing an object as a prop, and I already tried memo."

An honest note on code prompts. Polished has a guard: it checks that the rewrite keeps every number you said and stays close to your length, and if not, it types your own words (the raw transcript) instead. For prompts that ask for code, the guard sometimes returns your words unchanged rather than risk altering the request. When we said "ok so write me a um python function that takes a list of dates and uh returns the ones that fall on a weekend, and add tests, use pytest not unittest", Polished kept the original words, because the rewrite drifted. The agent still understands it, and "pytest not unittest" survives intact, which is the point.

Polished also suits non-native English speakers (it fixes grammar while keeping meaning), people who prefer to speak another language (it works in the language you speak, with an Always translate to English setting; on translations the guard checks numbers only, so check names), and older laptops, because nothing heavy runs locally. It needs an internet connection.

Privacy for developers

Here is where your speech goes in each mode:

  • Plain is local. Whisper runs on your CPU, audio never leaves the machine, and transcription works offline.
  • Polished sends the recording to AirTypes to transcribe and clean it up. Choose Plain for anything that should not leave your machine.
  • AirTypes always receives your account email and usage counts. In Plain, nothing else.
  • Your coding agent is separate. When you press Enter, Claude Code, Cursor or Copilot sends the prompt to its provider exactly as if you had typed it.

For the technical background, see how offline speech recognition works, and for a wider view, is voice dictation private?

Platforms and price

The AirTypes desktop app runs on macOS (Apple Silicon and Intel) and Linux today. The Windows desktop app is in development; you can join the Windows waitlist. On Windows today, you can Dictate in your browser (new in November 2026): sign in on airtypes.com, click the mic, speak, and the Polished text is ready to copy and paste into your terminal or editor. Nothing to install. It is Polished only, so it needs a connection and uses credit after the free first minute.

Plain is free with no time limit. Polished is prepaid credit with no subscription: $5 buys 600 minutes of speech, $10 buys 1,200, and credit never expires. When it runs out, Polished pauses and Plain keeps working. See pricing.

FAQ

Does AirTypes work in the terminal?

Yes. AirTypes types at your cursor in any desktop app, including terminal emulators. Click into the Claude Code prompt or any shell, hold the hotkey, speak and release. In the terminal, set the text output to Type at cursor, because many Linux terminals paste with Ctrl+Shift+V rather than Ctrl+V.

Can I dictate code with AirTypes?

You can, but it is awkward with any speech-to-text tool: brackets, operators and camelCase identifiers are slow to say and easy to mishear. Dictate intent in plain English, let your coding agent write the code, and type exact identifiers yourself.

Does AirTypes work with Claude Code, Cursor and GitHub Copilot Chat?

Yes, because AirTypes types at the cursor instead of plugging into any tool. Anywhere you can type a prompt, you can dictate one. There is no plugin to install, and AirTypes is not affiliated with Anthropic, Cursor, GitHub or OpenAI.

Does voice dictation send my code anywhere?

It depends on the mode. In Plain, Whisper transcribes on your CPU, no audio is uploaded, and AirTypes only receives your account email and usage counts. In Polished, the recording is sent to AirTypes to transcribe and clean it up. Either way, when you press Enter your coding agent sends the prompt to its provider, exactly as if you had typed it.

Should I use Plain or Polished for coding prompts?

Use Plain for anything about proprietary code, internal services or unreleased work, because the audio stays on your machine, and for short prompts you say cleanly. Use Polished for long, thinking-out-loud briefs you want tidied before you send them. Plain is free; Polished is prepaid credit.

Is AirTypes available on Windows?

The Windows desktop app is in development. Today you can use AirTypes on Windows in your browser: open Dictate on airtypes.com, click the mic and speak, and the Polished text is ready to copy and paste into your terminal or editor. Nothing to install. Plain, the free offline mode, needs the macOS or Linux app. Join the Windows waitlist on the download page and we email you once when the app ships.

Get AirTypes, free

Say the prompt, read it, press Enter. Plain is free and transcribes offline on your CPU; Polished is optional prepaid credit for tidying long briefs.

Download for macOS and Linux