Voice coding: the two kinds, the tools for each, and a setup that works

The short answer

Voice coding means writing software by speaking instead of typing. It comes in two kinds. In the first, you say the code itself through a command grammar: tools like Talon, Cursorless and Serenade turn set phrases into keystrokes, symbols and edits, so you can work with no keyboard at all. It takes practice, and its makers speak to people who cannot or should not type much. In the second kind, you describe the task in plain words to an AI agent, such as Claude Code, Cursor or GitHub Copilot, and the agent writes the code. All you need there is a way to turn speech into text: the voice input built into those tools, or a dictation app that types into any window. The second kind has no new vocabulary to learn, so it is the easier place to start. Either way, say the details out loud, type the exact names, and read the text before you press Enter.

Every fact about a tool on this page comes from that tool’s own website or documentation, read on 19 October 2026 and linked at the end. We have not tested these tools side by side, so nothing here is a ranking. We make DanDin, one of the dictation apps on this page, and we say so wherever it comes up.

What are the two kinds of voice coding?

They share a name, but they ask very different things of you.

The two kinds of voice coding. Command examples are from Talon's sample file (talonvoice.com, read 19 October 2026); the task example is ours.

	Saying the code	Describing the task
What you say	Short, fixed commands that the tool maps to keys and edits	Plain sentences, the way you would brief a colleague
Example	"select all", then "undo that"	"Add a dark mode switch to the settings page and keep the tests passing."
Who writes the code	You do, one command at a time	The AI agent does; you check its work
What you learn first	A vocabulary of commands	Nothing new; you learn to give clear tasks
Hands-free?	Can be: Talon also lists "Mouse by just looking"	Mostly not: you still press a key to start, and you review code
Tools	Talon, Cursorless, Serenade	Claude Code /voice, Cursor voice input, VS Code with Copilot, dictation apps

Swipe the table to see every column

Many people end up using both: the agent writes the big changes from a spoken task, and a few commands, or the keyboard, fix the small things.

Who does voice coding help?

People with hand, wrist or arm pain. The command tools speak to this directly. Serenade’s page says "Whether you have an injury or you’re looking to prevent one, Serenade can help you be just as productive without typing at all." The Talon Community Wiki, written by Talon’s users, says Talon "enables people who have limited use of their hands or want to stop using their hands to write code". If typing hurts, a tool can take load off your hands, but see a doctor about the pain as well.

People who think out loud. When you work with an AI agent, the quality of the result depends on how much you tell it. Anthropic’s best practices for Claude Code put it this way: "Claude can infer intent, but it can't read your mind." When you type, you write the short version. When you talk, you add the story: what you tried, which file you suspect, what must not break. The same guide suggests that for larger features you have Claude interview you and then write a spec. Answering questions out loud is easy.

People who think in another language. If you think in Arabic, or any language other than English, and write your code and prompts in English, voice lets you explain the idea in the language you think in, instead of translating it in your head while you type. Whether that works depends on the tool: Claude Code’s /voice, for example, lists 20 dictation languages and Arabic is not one of them.

Where voice is weak. Exact things: file paths, function names, symbols, error messages and code you want copied character for character. Paste or type those. Voice is also hard in a shared open office, and talking all day is tiring in its own way, so mix it with typing.

Kind 1: saying the code with a command grammar

A command grammar is a list of phrases, each tied to an action. You say the phrase, the action happens. It is precise and can replace the keyboard entirely, but you have to learn the phrases first.

Talon

Talon calls itself "Powerful hands-free input". Its download page offers macOS, Windows and Linux (X11). Its home page lists commands and dictation, "Mouse by just looking", and "Pop to click, hiss to scroll". It shows what a command file looks like:

# say it: do it
hello talon: "hello world"
select all: key(cmd-a)
undo that: key(cmd-z)

The words before the colon are what you say. The part after it is what Talon does. Real coding setups use large shared command sets plus your own additions. On price, the home page shows no figure; it links a "Talon+" tier on Patreon with "More features. Better support." The community wiki says "Talon is free to use and ongoing development is made possible by donations on Patreon."

Cursorless

Cursorless is a voice coding tool for VS Code that runs on top of Talon. Its install guide asks you to install Talon and a speech engine, the community Talon commands, VS Code, a VS Code Talon extension pack, the Cursorless extension, and the Talon side of Cursorless. Its docs sum up the idea: "Every Cursorless command consists of an action performed on a target." Its main GitHub repository had its last push on 6 October 2026.

Serenade

Serenade describes itself as "The open-source voice assistant for developers". Its page says it can run in the cloud or "completely locally", lists VS Code, IntelliJ, iTerm2, Chrome and Slack among the apps it works with, and offers a free download. One thing its page does not say: the project’s main GitHub repository had its last push on 11 June 2024. Check that it still runs on your setup before you depend on it.

Kind 2: describing the task to an AI agent

Here the agent writes the code, so voice only has to turn your words into text in the agent’s prompt. There are two ways to get that text.

Voice built into coding tools

• Claude Code. Run /voice, then hold Space while you talk. Anthropic’s docs say the audio goes to Anthropic’s servers, transcription does not count toward the limits in /usage, it needs a claude.ai sign-in, and it does not work in SSH or cloud sessions. It lists 20 languages, and Arabic is not one of them. Our guide on how to talk to Claude Code covers it in full.
• Cursor. Cursor 2.0, released on 29 October 2025, added Voice Mode: "Control Agent with your voice using built-in speech-to-text conversion." You can set custom submit keywords that start the agent. Cursor 3.1, on 13 April 2026, upgraded voice input in the Agents Window: "You can press and hold Ctrl+M to speak." The release notes do not list languages or say which plans include it.
• VS Code with GitHub Copilot. Microsoft’s docs say dictation works in "chat, the Agents window, editors, and terminals". In chat, select the microphone or press ⌘I (Ctrl+I on Windows and Linux), and "Dictation inserts text in the input but does not submit the request." On desktop, the speech model "processes microphone audio on your device". A separate Voice Mode, where the agent talks back, "is experimental and might change or be removed", and needs an eligible individual GitHub Copilot plan.

Built-in voice is the right choice when you work in one tool, your language is supported, and you like its defaults. Its limit is that it lives in that one tool: your commit message, the bug ticket and the chat with your team are back to typing.

System-wide dictation apps

A dictation app runs on your whole computer. You hold a key in any window, talk, and the text appears at your cursor, whether that is Cursor’s chat, Claude Code in a terminal, or a browser. Four that write about coding on their own sites, each in its own words:

• Wispr Flow: "Dictation built for developers". Its page says it "Handles camelCase, snake_case, and acronyms" and "Tags files in Cursor and Windsurf".
• Superwhisper: "Dictate your prompt and watch it land as text. Then send it to the AI agent that writes the code." It lists Claude Code, Codex, Cursor Composer, Copilot Chat in VS Code, Windsurf, Xcode and JetBrains, and says "It runs fully on-device if you want it to."
• Aqua Voice: "Talk your way through a task. Claude Code does the work." Its page says "The same key works over SSH, in a tmux pane, and in every other app on the machine."
• Willow Voice: "Speak rich prompts to Cursor, Claude Code, or Codex. More context in, better code out."
• DanDin (ours): hold a key, talk in Arabic, English or both, and clean text lands at your cursor. With Translate on, Arabic speech arrives as English.

Voice coding tools side by side

Every cell comes from the tool’s own page, read on 19 October 2026. "Not stated" means the page we read did not say. Prices are in US dollars, as each page showed them on that date.

Voice coding tools, from each maker's own page, read 19 October 2026. Not a ranking and not a test. DanDin is our product.

Tool	Kind	Where it works	Languages	Price (read 19 October 2026)
Talon	Command grammar	macOS, Windows, Linux (X11); the whole computer	Not stated on its home page	Free to use, per its community wiki; Talon+ through Patreon, no price shown
Cursorless	Command grammar	VS Code, with Talon	Not stated	Not stated
Serenade	Command grammar	VS Code, IntelliJ, iTerm2, Chrome, Slack and more; cloud or local	Not stated	Free download; last code push 11 June 2024
Claude Code /voice	Built into the agent	Claude Code CLI and its VS Code extension; not over SSH	20 listed; no Arabic	Not counted in /usage; needs a claude.ai sign-in
Cursor voice input	Built into the agent	Cursor’s Agent; hold Ctrl+M in the Agents Window	Not stated in the release notes	Not stated in the release notes
VS Code dictation	Built into the editor	Chat, Agents window, editors, terminals; on device on desktop	"auto" follows your system language; no list	Dictation built in; Voice Mode needs an individual Copilot plan
Wispr Flow	Dictation app	Any app; Mac, Windows, iPhone, Android	100+	Free plan; Pro USD 15 a user a month, or USD 12 a month billed yearly
Superwhisper	Dictation app	Any app; macOS, Windows, iOS, Android; on-device option	Not stated on its voice coding page	Free tier "that doesn’t expire"; paid price not read (its pricing page returned 404)
Aqua Voice	Dictation app	Any app; Mac, Windows, iPhone, Android	49	Free for 1,000 lifetime words; Pro USD 8 a month billed yearly
Willow Voice	Dictation app	Any text field; Mac, Windows, iPhone	100+	Free basic plan; Pro USD 12 a month billed yearly
DanDin (ours)	Dictation app	Any app with a text box; macOS 12 or newer, 64-bit Windows 10 and 11; no Linux, no phone app	Arabic and English, alone or mixed	Free for 2,000 words a week; Pro USD 16.99 a month or USD 149 a year

Swipe the table to see every column

Two things the table cannot show. First, a language count says nothing about how well a tool handles your accent or a sentence that mixes two languages, so try your own speech before you pay. Second, prices change; check the maker’s page on the day you buy.

A practical voice coding setup

These steps work with any of the tools above.

1. Pick your kind

   If typing hurts or you want to leave the keyboard entirely, start with a command grammar (Talon, then Cursorless) and expect to practise. If you already work with an AI agent, start with the second kind. You can add the other later.

2. Pick where the words come from

   Use your agent’s built-in voice if you work in that one tool and it lists your language. Use a dictation app if you switch between tools, speak a language the built-in voice lacks, or run Claude Code over SSH, where Anthropic says /voice does not work.

3. Set a start key you can hold without looking

   Give the app microphone access. On a Mac, a dictation app also needs Accessibility so it can paste into other apps. Test it once in your editor, once in your terminal and once in your agent’s chat.

4. Say the task in parts

   The goal, where in the code, the rules, and how to check it is done. Our Claude Code guide walks through the four parts with an example, and they work in Cursor and Copilot too.

5. Type the exact bits

   After talking, paste the error message, type the file path, or pick the file from your agent’s file picker. Mixing voice and typing in one message is normal.

6. Read, then press Enter

   Know what sends the prompt. Claude Code’s hold mode waits for Enter unless you turn on autoSubmit, and its tap mode sends by itself once the text is three words or more. Cursor can start the agent on a submit keyword you choose. VS Code’s chat dictation does not send. DanDin never presses Enter.

A worked example: rambling in, clear request out

Spoken tasks are messy. People change their mind halfway through a sentence. This is a request in English, written as said and run through DanDin’s engine on 11 October 2026, in its "What I meant" mode, with Cursor as the app in front. It is the same example as on our homepage.

As said: ok so I want a booking page, like with name and phone, no wait, and email too, and the customer picks the day and time, and after booking they get a confirmation email, I mean a text, not an email.

DanDin’s real output: I need a booking page that collects the customer's name, phone number, and email address. The page should allow users to select a date and time for their appointment. Once the booking is complete, the system should send a confirmation via text message.

Look at the two corrections. "No wait, and email too" became a third field. "A confirmation email, I mean a text, not an email" became a text message, and the confirmation email is gone. A raw transcript would have handed the agent both versions and left it to guess.

What the output still lacks is what only you know: which folder the pages live in, what the project already uses for forms, and how to check the page works. Say that in a second breath, or type it, and the task is ready for the agent.

Where DanDin fits

DanDin is our product, so weigh this part with that in mind. DanDin (dandin.com) is a dictation app for Mac and Windows, built for Arabic and English. It is the second kind of voice coding: you describe, the agent codes.

• Arabic or mixed speech, English for the agent. Speak Arabic, English or both in one sentence. Turn on Translate to English and Arabic speech arrives in Cursor or Claude Code as English. Our page on Arabic and English in one dictation explains how the mixing works.
• Cleaned text. It drops the "ums", applies your mid-sentence corrections and adds punctuation, as in the example above.
• Into any window. Hold Fn on a Mac, or Ctrl and the Windows key on a PC, talk, and let go. DanDin pastes at your cursor and puts back whatever you had copied before, whenever that is safe. It never presses Enter, so you read the task first. On a Mac, Secure Keyboard Entry blocks it, as above.
• The limits. It needs the internet. One dictation runs up to 30 minutes. There is no Linux version and no phone app. It is not a command grammar: it cannot select a line or run an editor command by voice. If you cannot use your hands at all, Talon is the tool built for that.

The per-app pages show the details for DanDin with Claude Code and DanDin with Cursor. If you are weighing dictation apps against each other, our Wispr Flow alternatives page compares them. To try the idea with nothing to install, our free voice to prompt tool turns a spoken idea into a clear English prompt in your browser.

If you build whole apps by describing them, our guide to vibe coding with voice covers the workflow and the risks. And once your spoken tasks keep repeating the same rules, put those rules in a skill: our Claude Skills guide has a complete one to copy.