Voice Input for Developers (2026): Claude Code, Cursor, VS Code and the Terminal

Before you install anything: the tool you already use probably has dictation in it. Claude Code shipped /voice in the spring, Cursor has had a microphone in the chat box since October 2025, and VS Code has dictated into the editor for longer than that. Everything on this page was read off official documentation, changelogs and package registries on 27 August 2026, including the conditions that the guides written by dictation vendors leave out, and there are several of those.

By the VoiceBoard editorial teamUpdated August 28, 2026

Two different jobs live under the phrase "voice coding"

Almost every argument about this topic comes from mixing them up.

Dictating intent to an agent. You say what you want, prose goes into a prompt box, the model writes the code. Nothing about your editor changes. This is what the built-in features do, this is what a dictation app does, and for most people reading this it is the whole task. Setup time: about a minute.

Driving the editor by voice. You issue structured commands that move the cursor, select a token, delete an argument, rename a symbol. No typing anywhere in the loop. This is Talon and Cursorless, it is a skill you learn rather than an app you install, and the people who master it mostly did so because typing started hurting.

The first job got easy in 2026. The second has been possible since roughly 2018 and is still hard. If you came here because your wrists hurt, skip to the Talon section, because a dictation app will not solve your problem.

Why this suddenly makes sense

A prompt to a coding agent is not a line of code. It is two hundred words of prose describing what you want, which files are involved, what you already tried and what the constraints are. Prose is the exact thing speech is good at and the exact thing that is slow to type.

The numbers people quote for this come from one 2016 study by Ruan, Wobbrock, Liou, Ng and Landay (arXiv:1608.07323). On an iPhone 6 Plus, English speech input ran at 153 words per minute against 52 for the on-screen keyboard, roughly three times faster, with Mandarin at 123 against 43. Worth knowing before you repeat the figure: that was thumb typing on a phone. The comparison you care about is a physical keyboard, and the largest study of those, Dhakal and colleagues at CHI 2018 with 168,000 participants and 136 million keystrokes, also landed on an average of 52 words per minute, with the top five percent above 80.

So for an average typist the gap really is around three to one. If you type at 100 words per minute the gap is closer to 1.5, and the case for voice stops being about speed. The same 2016 study found speech left slightly more uncorrected errors in the final text, 1.30 percent against 0.79 percent, which matters more when you are dictating a file path than when you are dictating a sentence.

The honest version of the pitch: dictation does not make you a faster programmer. It makes long prompts cheap, so you write the three-sentence version instead of the six-word version, and the agent guesses less.

At a glance

Claude Code: the /voice command

This is the one the search traffic is about, and it is also the one described least accurately in the articles ranking for it, so here is the full picture from Anthropic's documentation.

Turning it on. Type /voice in the CLI. The first run does a microphone check, and on macOS that is when your terminal finally asks for microphone permission. The prompt footer confirms the state:

Voice mode enabled (hold). Hold space to record. Dictation language: en (/config to change).

/voice hold, /voice tap and /voice off set the mode directly. The setting survives restarts, and you can skip the command entirely by putting it in your user settings file:

{ "voice": { "enabled": true, "mode": "tap" } }

Hold mode, the default, is push-to-talk on the space bar. There is a quirk worth knowing: Claude Code detects a held key by watching for key-repeat events from the terminal, so there is a short warmup, the footer says keep holding… before it says listening…, and the first stray space characters get cleaned up automatically. Text appears dimmed as you speak and is inserted at the cursor when you release, which means you can type half a sentence, speak the rest, and keep going.

Tap mode removes the warmup: tap space once to start, tap again to stop, and the prompt submits itself if the transcript is at least three words. Recording stops on its own after 15 seconds of silence or two minutes total. The first tap only works when the input is empty, so you can still type a space normally.

Two settings worth changing on day one. Add "autoSubmit": true to the voice object if you want hold mode to send on release rather than waiting for Enter. And if the space bar warmup annoys you, rebind the action in ~/.claude/keybindings.json, because a modifier combination starts recording on the first keypress with no warmup at all:

{ "bindings": [ { "context": "Chat", "bindings": { "meta+k": "voice:pushToTalk" } } ] }

The transcriber has been pointed at our vocabulary rather than at general speech. Anthropic says it handles terms like regex, OAuth, JSON and localhost correctly, and that your current project name and git branch name are fed in as recognition hints automatically. That last detail is the sort of thing that decides whether dictation is usable, and no third-party tool can do it without reading your repo.

Twenty dictation languages, set through /config or the language setting: Czech, Danish, Dutch, English, French, German, Greek, Hindi, Indonesian, Italian, Japanese, Korean, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Turkish, Ukrainian. Anything else falls back to English with a warning on enable.

Now the conditions, which is where the vendor guides go quiet. Voice dictation requires:

• A Claude.ai account. It is unavailable if Claude Code is pointed at an Anthropic API key directly, at Amazon Bedrock, at Google Cloud's Agent Platform or at Microsoft Foundry. A large share of professional users run exactly those setups, and for them /voice simply is not there.
• A local microphone. It does not work over SSH or in Claude Code on the web, because the microphone is on your laptop and the process is not.
• WSLg, if you run Claude Code under WSL. That comes with WSL2 from the Microsoft Store. On WSL1 you are told to run Claude Code in native Windows instead.

Audio goes to Anthropic's servers. The documentation puts it in four words: "Audio is not processed locally." On the other hand, transcription "does not consume Claude messages or tokens and does not count toward the limits shown in /usage", so the feature is free in the sense that matters.

The VS Code extension has the same feature and the same Claude.ai requirement, and it is unavailable in VS Code Remote sessions, meaning SSH, Dev Containers and Codespaces. Your organisation's administrator can also switch the whole thing off by policy, in which case /voice tells you so.

On Linux, recording uses a native module, and if that fails to load Claude Code falls back to arecord from ALSA utils or rec from SoX. Under WSLg you need the PulseAudio backend explicitly, sudo apt install sox libsox-fmt-pulse, because plain sox pulls the ALSA backend and there is no /dev/snd to record from. That single line will save an evening for anyone who develops in WSL.

When it arrived. Anthropic does not date its changelog, so we went at it sideways. Across all 377 versions in the public changelog, the first entry mentioning voice is 2.1.69, which added ten more dictation languages for a total of twenty, and the npm registry gives that version a publish date of 4 March 2026. The feature therefore landed at the very start of March 2026 and has been getting fixes at a steady clip since, most of them about Linux audio, WSL and terminals that do not send key-repeat events.

Cursor: the microphone in the chat box

Cursor added voice in Cursor 2.0 on 29 October 2025, described in its changelog as controlling the agent with your voice through built-in speech to text, with custom submit keywords you define in settings so that saying your trigger word starts the run. Cursor 3.1, on 13 April 2026, rebuilt it: hold Ctrl+M to speak, and the editor now records the full clip and transcribes it in one batch rather than streaming, which the changelog says improved quality. While recording you get a waveform, a timer and cancel or confirm buttons.

One page currently ranking for this query states that Cursor's voice input "is not covered by public documentation we could verify". It is. The changelog entries above are public, dated and specific, which tells you something about how carefully this field is written.

What Cursor's microphone does not do is leave the chat box. It does not reach Cmd+K, the editor itself, the integrated terminal, or the browser tab where you are reading an API doc. If you run Claude Code or Codex inside Cursor's terminal, the chat microphone is no help there at all, and that is the single most common reason developers using Cursor end up installing a system-wide tool anyway.

Cursor's community forum also carries a steady stream of voice bug reports: input that starts an audio context but captures silence, waveforms that animate while no text ever appears, and dictated text that lands at the end of the first line regardless of where the caret is. None of that means the feature is broken for everyone. It does mean you should test it before you build a habit around it.

VS Code and Copilot Chat

VS Code dictates in two places, and the shortcuts are worth committing to memory: Ctrl+I (⌘I on Mac) speaks into chat, Ctrl+Alt+V (⌥⌘V) dictates into the editor at the cursor. Both are press and hold. There is also "Hey Code" keyword activation if you would rather not touch a key at all.

The part that matters for anyone with a strict employer: Microsoft's documentation states that the default model "processes microphone audio on your device" and that after the initial model download, speech recognition does not require an internet connection. This is a genuinely capable, genuinely local option that ships from the same vendor as your editor, and it is free.

On mainstream desktop platforms the capability is now built in. The separate VS Code Speech extension is still needed on the web, on Intel Macs, on 32-bit and Arm32 systems and on musl-based Linux distributions. It sits at 1,417,222 installs, version 0.16.0, and covers 26 languages through the accessibility.voice.speechLanguage setting.

The limit is the obvious one. It works inside VS Code and nowhere else. And because Cursor is a VS Code fork rather than VS Code, do not assume the extension behaves identically there.

Terminals

Warp is the outlier: it has voice as a first-class feature, configured under Settings, Agents, Warp Agent, Voice, usable across its input surfaces rather than only in agent mode. Warp's own documentation says transcription "is powered by Wispr Flow" and that audio is processed in real time and not kept as a recording afterwards. If you already pay for Wispr Flow, you are running the same engine in two places.

For iTerm2, Windows Terminal, Ghostty, Alacritty or a plain gnome-terminal, there is no built-in dictation and there is not going to be. A terminal is a text surface like any other, so the answer is a system-wide layer that types into the focused window. This is also the answer for JetBrains IDEs, for Zed, for Neovim, and for the CLI agents you run inside any of them.

The system-wide layer

These sit in the background, listen while you hold a hotkey, and paste finished text wherever the cursor is. One tool, every window. They are compared on the same terms here: what you pay, where the audio goes, where it runs.

Wispr Flow

The one built specifically for this audience. Its developer page promises dictation that "understands dev jargon", handles camelCase, snake_case and acronyms, picks up product names like Supabase or MongoDB, and tags files in Cursor and Windsurf so the right context lands in your prompt. It claims 4x faster shipping, which is the vendor's number and not a measurement anyone has published.

Here is what earns it a recommendation despite the price. You hold a key, you talk the way you talk, and clean prose lands in the window you were in: filler gone, punctuation placed, the sentence you restarted rendered as the sentence you meant. It works in the terminal, in the chat panel, in a GitHub comment and in Slack without knowing or caring which is which, so the mental overhead of "which microphone do I press here" disappears after two days. Warp's team picked it as the engine behind their own voice feature, which is a real vote from people who evaluated the alternatives.

The fine print. Free tier is 2,000 words a week on desktop, Pro is $15 a month or $12 a month billed yearly. It is a cloud service, so no connection means no dictation. Windows is x64 only: ARM machines, meaning the Snapdragon laptops, are not supported, and neither are virtual machines or remote desktop, which rules it out for a chunk of enterprise setups. There is no Linux build. Our full review is here, and the alternatives are here.

Try Wispr Flow

Superwhisper

The news is a day old at the time of writing: on 26 August 2026 Superwhisper launched a permanent free tier. Zero dollars now buys voice to text in any app, meeting recording and transcription, support for 100+ languages and unlimited use of the Whisper models. Pro, at $8.49 a month, adds your own API keys, cloud and local AI models, translation, and transcription of audio and video files. New users get 3,000 words of Pro to try, after which the free features stay yours indefinitely.

It runs on Mac, Windows and iOS, and it is one of the few commercial apps that will process your speech entirely on your own machine. Offline models want Apple Silicon; the vendor says Intel Macs are better off on cloud models.

There is also a dedicated Claude Code integration, added in April 2026, which puts a floating Superwhisper window in front of the agent so you can talk to it without switching to the terminal. Setup is a one-line install script from their site. Equivalent pages exist for OpenCode, Codex, Pi and Grok CLI.

Handy

Free, MIT licensed, and it does one thing: press a shortcut, speak, get text in the focused field, with nothing leaving the computer. Transcription runs on Whisper models or on Parakeet V3, which is CPU-optimised and detects language automatically, with Silero voice activity detection filtering the silence. Windows, macOS and Linux, current release v0.9.6 from 24 August 2026, and 30,430 stars on GitHub.

The project describes itself as trying to be the most forkable speech-to-text app rather than the best one, which is an unusually honest positioning and also the right expectation to have. There is no polish layer that cleans up your grammar, no dictionary that learns your colleagues' names. For a developer who is comfortable picking a model size and living with occasional rough edges, Handy is the strongest free option on this page.

VoiceInk

macOS on Apple Silicon only, local models on the Neural Engine, open source, 6,123 stars, and priced as a one-time purchase: $29 for one Mac, $49 for two, $69 for three, with lifetime updates. Cloud providers are available only if you supply your own API key and choose to.

If your Mac is the machine with the client code on it, this is the shortest route to dictation that never sends audio anywhere, and paying $29 once against $144 a year for a subscription pays for itself in about ten weeks. Our Mac roundup goes deeper.

See VoiceInk

Voicy

The commercial option that actually ships a Linux build: .deb, .rpm and AppImage, covering Ubuntu, Debian, Fedora and Arch. It writes into every app, terminal and browser on the system, works in VS Code and JetBrains editors, and also runs on Windows, macOS, Chrome and mobile. Free trial is 30 minutes of recording, Pro is $8.49 a month, and there is a lifetime licence at $260. Students and people with a disability get 20 percent off.

The honest limit: there is no offline mode. Recognition runs through Groq's infrastructure, so audio leaves your machine every time. Voicy says it is deleted right after processing and never used for training, which is a reasonable policy and still not the same as local processing.

See Voicy

Voibe

Worth naming for one feature the others do not have: a developer mode that resolves spoken identifiers against your actual workspace, so "auth middleware dot t s" comes out as authMiddleware.ts rather than as words. That is the single most annoying failure mode of dictating in an editor, solved directly. $7.50 a month, $59 a year, or $149 once, on Mac and Windows.

One correction that gets repeated wrongly elsewhere: Voibe is fully on-device on Apple Silicon, but on Windows it runs in zero-retention cloud mode, per its own FAQ. Zero retention is a meaningful policy. It is not the same as the audio never leaving the room.

If the code is not allowed to leave the machine

This is the question that separates a hobby setup from one you can use at work, and the marketing copy in this category muddies it deliberately. Two things get conflated: where the audio is processed, and what the vendor promises to do with it afterwards. Only the first one is a technical fact.

Audio never leaves your computer: VS Code Speech, Handy, VoiceInk on Apple Silicon, Superwhisper on local models, Voibe on Apple Silicon, Talon.

Audio goes to a server: Claude Code /voice, Cursor voice, Warp voice, Wispr Flow, Voicy, Voibe on Windows, Superwhisper on cloud models.

Note where Claude Code sits, and note the second-order effect: /voice is unavailable on Bedrock, on Google Cloud's Agent Platform and on Microsoft Foundry, which are the exact deployments regulated companies choose. If your employer put Claude Code behind Bedrock for compliance reasons, the built-in microphone is not an option you have to argue about, because it will not appear.

For that reader the recommendation costs nothing: Handy if you want cross-platform and free, VS Code Speech if you live in one editor, VoiceInk if you are on an Apple Silicon Mac and would rather pay once for something more finished. Check with whoever owns your security policy anyway, because "runs locally" is a claim you should verify rather than inherit from a page like this one.

Talon and Cursorless: the other sport

If your hands are the reason you are here, none of the above is your answer. Dictation types words. It does not move a cursor, select the third argument, or refactor anything.

Talon is a hands-free input system: voice commands, noise control (a hiss or a pop becomes a click), eye tracking, and Python scripts to define whatever you want. It runs offline on its own speech engine, on macOS, Windows and Linux under X11, and it is free, with a Patreon tier for beta builds. Look at the version numbers before you commit: the public release is 0.4.0, dated 24 July 2023 in Talon's own changelog. The beta channel is where the recent work sits, including a hybrid Whisper engine, a "mixed mode" that accepts commands and dictation simultaneously, and the Conformer D model that the changelog credits with about 20 percent better accuracy. Development is alive; the stable download just does not show it.

Cursorless is the layer that makes editing structural rather than positional. It paints a coloured "hat" over one character of every visible token, and you address code by those hats: "chuck bat" deletes the token whose hat sits on that b. It is a VS Code extension driven by Talon, MIT licensed, 1,335 stars, and it was committed to as recently as 26 August 2026. The community command set at talonhub/community, which is what most people actually run on top of Talon, was last touched on 24 August 2026. Both projects are healthy.

Do not let anyone tell you this is a quick win. You are learning a spoken language for editing text, and the payoff arrives in weeks rather than hours. The most-read account of it, Xe Iaso's Cursorless is alien magic from the future, reached the front page of Hacker News on 10 November 2023 with 790 points and 337 comments, and reading that thread is the fastest way to calibrate: the conversation is almost entirely about RSI, tendinitis and staying employable, and barely at all about speed. That is who this is for.

Serenade still appears in lists of voice coding tools. Its site is up and its download button works, but the GitHub repository has had no push since 11 June 2024. Treat it as unmaintained until that changes.

How to dictate so the output is usable

The habits below come from the shape of the problem rather than from any tool, and they apply equally to /voice, to Cursor and to a system-wide app.

Speak intent, not syntax. "Paginate the users table, twenty per page, server side" beats narrating a for loop. The agent is better at syntax than your microphone is, and every symbol you try to pronounce is a chance to be wrong.

Ten to thirty words per burst, then pause. Long unbroken monologues degrade, and short ones lose the context that makes the transcript resolve correctly.

Let the tool resolve names, and check the ones it cannot. Claude Code feeds your project and branch names to the transcriber; Voibe matches spoken identifiers against the workspace; Wispr Flow learns jargon as you use it. None of them will reliably get a colleague's surname or an internal service codename, so glance at the prompt before you send it.

Hold off on auto-submit until you trust what comes back. Sending automatically at three words is convenient right up to the moment a half-heard sentence starts an agent run you did not intend, and you spend the next two minutes stopping it.

Finally, keep typing available. Every good implementation lets you mix the two inside one message: file paths, flags and regexes are faster typed, and there is no prize for purity.

How we checked this, and what we did not check

No accuracy benchmark was run for this article, and we are not going to imply otherwise. Read the pages currently ranking for these queries and you will find confident figures, "around 98% accuracy", "200ms versus 700ms", "97.3% on the AISpeak benchmark", published by companies selling the tool that finished first in their own comparison, with no recordings, no scripts and no raw data attached.

What we did instead, on 27 August 2026, was verify the checkable things at the source.

• Read Anthropic's voice dictation documentation in full, which is where every condition in the Claude Code section comes from.
• Searched all 377 versions in the public Claude Code changelog for voice entries, and dated version 2.1.69 through the npm registry's publish timestamp.
• Read the Cursor 2.0 and 3.1 changelog entries for the dates, the hotkey and the switch to batch transcription.
• Read Microsoft's accessibility documentation and the marketplace listing for the VS Code Speech extension.
• Read Warp's own documentation for the confirmation that its transcription runs on Wispr Flow.
• Read Talon's public changelog for the release date and the beta feature list.
• Queried the GitHub API for the star counts and last-commit dates of Cursorless, the Talon community command set, Handy, VoiceInk and Serenade.
• Read the current pricing pages of Wispr Flow, Superwhisper, Voicy, Voibe and VoiceInk rather than copying prices out of other roundups.

Three things we could not verify and are flagging rather than glossing over. Vendor accuracy and latency claims are marketing, reported here as such and not endorsed. Cursor does not document where its transcription happens or who provides it, so the row in our table says "not documented" rather than guessing. And Reddit is not reachable from our research environment, so the community sentiment referenced here comes from sources you can click: a Hacker News thread with a date, and Cursor's own public forum.

Tools we looked at and left out

Aqua Voice, Willow Voice, Spokenly, Vibe Typer, VoiceDash and the rest of the blog-first crowd. These are real products with real users, and their content operation is the reason this topic is hard to research: each one publishes a guide that treats the built-in features as a problem and its own download as the answer. Some are surely good. We have not tested them and will not rank what we have not checked.

MCP voice servers such as mcp-voice-hooks. Clever, and they let an agent speak back to you rather than only listen. They are also an extra moving part in your agent config for a job that a hotkey already does. Worth a look if you want two-way conversation, not if you want faster prompts.

Ski, freeflow, Yap and the weekly Show HN wave. New local dictation tools launch on Hacker News at a rate of roughly one a week right now, most of them free and offline. If you like being early, that feed is the place to watch. Handy is the one that has already accumulated the users, the releases and the packaging.

Dragon Professional. Still the benchmark for correcting and formatting by voice, still Windows only, still around $699 at resellers, and its consumer distribution has been dismantled: what happened to Dragon. It was never built for editors or terminals.

System dictation on macOS and Windows. Free, already installed, fine for a sentence in a text field. Both fall apart on technical vocabulary, which is the entire content of a developer's speech.

What to do next

Spend five minutes on what you already have: /voice in Claude Code, Ctrl+M in Cursor, Ctrl+I in VS Code. If the only thing that annoys you is having a different microphone in every window, that is exactly the problem a system-wide layer solves, and you can settle it on the free tier of Wispr Flow before you pay anything. If your code cannot leave the machine, install Handy instead, or VoiceInk if you are on an Apple Silicon Mac and want something more finished for $29 once. And if you are here because typing hurts, none of that is your answer: start with Talon and Cursorless, and give it weeks.