Voice Typing in VS Code, Cursor, and Claude Code
If you write software, a surprising amount of your day isn’t really code — it’s prose. Comments, docstrings, commit messages, PR descriptions, and increasingly the natural-language prompts you feed to AI coding agents. Voice typing handles all of that well, and HyperVoice does it in VS Code, Cursor, and Claude Code without any editor extension because it injects text at your cursor in whatever app is focused.
This post is specifically about editors, terminals, and AI agents. For the broader case — Slack, email, tickets, the whole developer workflow — see HyperVoice for Developers. Here we’ll go deep on the parts that are unique to writing inside an IDE and talking to coding agents.
What Voice Is (and Isn’t) Good For in Coding
Let’s be honest up front: dictation is bad at raw code. Brackets, semicolons, camelCase and snake_case identifiers, operators, nested generics — none of that maps cleanly to spoken language, and forcing it through your voice is slower and more frustrating than just typing. Anyone who tells you they dictate their actual source line by line is either doing something very unusual or exaggerating.
Where voice wins is everything in your codebase that reads like English:
- Comments and docstrings — full sentences explaining intent.
- Commit messages and PR descriptions — prose summaries of what changed and why.
- READMEs, design docs, and ADRs — long-form writing that lives next to the code.
- Natural-language prompts to AI agents — the big one, covered next.
The rule of thumb: if what you’re about to type would read naturally out loud as a sentence, dictate it. If it’s a token stream of symbols and identifiers, type it. You’ll keep one hand near the keyboard for the code and offload the prose to your voice. That split is the whole game.
Dictating to AI Coding Agents Is the Killer Use
This is where voice typing stops being a nice-to-have and becomes genuinely faster than the keyboard. Tools like Claude Code (running in your terminal), Cursor’s AI chat, and GitHub Copilot Chat all take the same kind of input: a natural-language description of what you want. That’s prose, and prose is exactly what dictation is good at.
The prompts you give an AI agent are often long. You’re explaining context, constraints, edge cases, the shape of the output you want, what to avoid. Typing a paragraph like that is slow and you tend to truncate it — you write a terse prompt because the full one is too much effort, and you get a worse result. Speaking it removes that friction. You ramble out the full context in fifteen seconds, the agent gets everything it needs, and the answer is better because the prompt was better.
A typical loop with Claude Code in a terminal:
- Focus the terminal where Claude Code is waiting for input.
- Press the global hotkey (default Ctrl + Shift + Space).
- Say something like: “Refactor the auth middleware so it reads the token from the Authorization header instead of the cookie, keep the existing error responses, and add a test for the missing-header case.”
- Stop recording. The full prompt lands at the cursor. Hit Enter.
The same flow works in Cursor’s chat panel or Copilot Chat — click the input, hold or tap the hotkey, speak the request. Because HyperVoice pastes at the cursor of the focused app, it doesn’t care whether that cursor is in a terminal, a chat sidebar, or a text box on a web page. There’s nothing agent-specific to configure.
If you want your spoken prompts tightened up before they’re sent, turn on a cleanup processing mode (more on that below). It strips filler words and false starts so the agent sees a clean instruction instead of “um, okay so, like, can you maybe…”.
Comments, Docstrings, and Commit Messages
The other everyday win is the prose that lives inside your code and your git history.
Comments and docstrings. Put your cursor where the comment goes, hit the hotkey, and explain the function the way you’d explain it to a teammate. You’ll usually type the comment delimiters yourself (//, #, """) and dictate the sentence into them — which is the right division of labor, since the delimiters are syntax and the explanation is prose. A custom processing mode like “Format as a JSDoc block” can wrap the output for you if you want to skip even that.
Commit messages. In your terminal, type git commit -m ", hit the hotkey, and describe the change. Stop recording and the message appears between the quotes. If you care about format, set up a custom mode with a prompt such as: “Rewrite as a conventional commit — type prefix (feat, fix, refactor, docs, chore), subject under 72 characters, optional body.” You speak a loose description; the mode returns a clean message.
PR descriptions. Click into the GitHub or GitLab description field and talk through what the PR does, why, and anything reviewers should look at. This is long-form prose that most people under-write because typing it is tedious — speaking it tends to produce more thorough descriptions, not fewer.
Setup in VS Code, Cursor, or a Terminal
There’s no setup specific to any of these tools, and that’s the point. HyperVoice is a system-wide app, not a VS Code or Cursor extension — it listens for a global hotkey and injects the transcribed text at your cursor in whichever window is focused. So “setting it up for VS Code” is really just:
- Install HyperVoice (Windows 10+, Linux x64 in beta, or macOS on Apple Silicon in beta) and let it download a model on first run.
- Confirm the global hotkey in Settings — the default is Ctrl + Shift + Space.
- Focus your editor or terminal, press the hotkey, speak, and the text appears at the cursor.
Because the hotkey is global, the exact same flow works in VS Code, Cursor, Windows Terminal, PowerShell, Vim in insert mode, a JetBrains IDE, or a browser tab. No plugin to install per editor, nothing to keep in sync with editor updates. (iOS is on the roadmap, not available today.)
Tips for Coding Workflows
A few configuration choices that make voice typing noticeably better for developers:
- Build a custom dictionary for your jargon. Whisper does a good job on plain English but will mangle library names, product names, and acronyms. Add dictionary replacements so “next js” becomes “Next.js”, “kubernetes” stays “Kubernetes”, and your internal service names come out right every time. This is the single highest-leverage tweak for technical dictation.
- Use push-to-talk for short bursts. In Settings, switch from toggle to push-to-talk. Hold the key while you speak a commit message or a quick prompt, release to transcribe. It pairs well with a side-mouse button or a footpedal and avoids leaving the mic open.
- Turn on a cleanup mode for prose. A processing mode like Clean Up removes filler and fixes punctuation so dictated comments and PR descriptions read cleanly. Note that cleanup modes run in the cloud — via HyperVoice Cloud (a Pro add-on) or your own OpenAI/Anthropic key (BYOK, available on every tier). Transcription itself stays local; only the transcribed text is sent for cleanup. Leave cleanup off and nothing leaves your machine at all.
- Pick a model that keeps up. If transcription feels slow you’ll drift back to typing. The larger Whisper models and Parakeet are more accurate; a GPU (Vulkan, with CPU fallback) keeps them fast. Match the model to your hardware so dictation feels instant.
A Note on Wrists
Developers are squarely in the population that ends up with RSI and wrist pain — the keystroke counts are brutal. Moving comments, commit messages, PR descriptions, and AI prompts to voice is an easy way to shave a meaningful chunk off your daily typing without giving up the precision typing you need for actual code. If that’s on your radar, we wrote a fuller guide: Voice Dictation for RSI and Wrist Pain. The same prose-vs-syntax split applies — voice for the sentences, keyboard for the symbols, fewer keystrokes overall.
Getting Started
You can try all of this on the free tier — 500 words a day, no card required. Install HyperVoice, open your editor or a terminal, and dictate your next commit message or your next prompt to Claude Code. Sign up free and see how much of your coding day is actually prose.
If you want the wider picture beyond the editor — tickets, bug reports, chat, email — start with HyperVoice for Developers or Filing Bug Reports and Tickets with Voice, and the homepage has the full feature rundown.
Frequently asked questions
Can I dictate code with HyperVoice?
You can dictate the prose parts of coding well — comments, docstrings, commit messages, PR descriptions, and especially natural-language prompts to AI coding agents. Dictation is not good for raw code syntax like brackets, camelCase identifiers, and operators; type those by hand. The sweet spot is anything that reads like a sentence rather than a token stream.
Does HyperVoice work as a VS Code or Cursor extension?
No. HyperVoice is a system-wide desktop app, not an editor extension. It injects transcribed text at the cursor of whatever app is focused, so it works in VS Code, Cursor, JetBrains IDEs, terminals, and browsers without installing anything inside the editor. The global hotkey is what makes it work everywhere.
Is voice transcription private when coding?
Speech-to-text runs 100% locally using Whisper or NVIDIA Parakeet, so your audio never leaves your machine for transcription. Optional AI cleanup modes run in the cloud (HyperVoice Cloud or your own OpenAI/Anthropic key) and send only the transcribed text, never the audio. If you leave cleanup off, nothing leaves the device.
Related posts
Filing Bug Reports and Tickets with Your Voice
Stop context-switching to write up tickets. HyperVoice's Ticket / Issue mode turns your spoken description into a structured bug report or task — ready to paste into Jira, Linear, or GitHub.
HyperVoice for Developers: Voice Dictation in VS Code, Cursor, and the Terminal
How developers use HyperVoice to dictate code comments, commit messages, documentation, and chat with AI coding agents — all without leaving their editor.
Voice Dictation on Linux: A Practical Guide
Good native Linux voice dictation is genuinely hard to find. Here's an honest guide to your options, plus how HyperVoice does local Whisper speech-to-text on Linux.