1. Five tests for a Claude Code dictation tool
A dictation app that is excellent in email can still be awkward in Claude Code, because the prompt lives in a terminal and the text you speak is about code. Judge each option on five things: where the speech is turned into text, which systems it runs on (including Linux, WSL and SSH sessions), how it charges, what it does in a terminal, and whether it spells your project’s names.
Terminal behavior is the one people forget. Claude Code’s hold mode inserts the transcript and waits for Enter, while its tap mode submits automatically once a transcript has at least three words. External apps paste or type into the terminal, and some also press Enter or reformat text; Wispr Flow’s own documentation warns that its context-aware capitalization, spacing and punctuation can transform shell commands. Prefer a tool that leaves the prompt for you to read before Claude Code starts working.
- Processing: on your computer, or on a vendor’s servers.
- Reach: macOS, Windows, Linux, WSL, and a terminal connected over SSH.
- Price: free, subscription, word cap, or one-time license.
- Terminal: paste or type, auto-formatting, and whether Enter is pressed for you.
- Names: whether file, function and variable names come out as the code spells them.
2. Claude Code /voice: the built-in choice and its three limits
Run /voice and hold Space, or run /voice tap to start and stop with single taps. Anthropic says transcription is tuned for coding vocabulary, recognizes terms such as regex, OAuth, JSON and localhost, and adds your project name and Git branch name as hints automatically. It supports 20 languages, follows Claude Code’s language setting and falls back to English. Transcription does not consume Claude messages or tokens or count toward /usage limits. Claude Code itself requires a Pro, Max, Team, Enterprise or Console account, and the free Claude.ai plan does not include it.
The limits are firm. First, audio is streamed to Anthropic’s servers; nothing is processed locally. Second, /voice needs a Claude.ai sign-in, so it is unavailable with an Anthropic API key used directly, Amazon Bedrock, Google Cloud’s Agent Platform or Microsoft Foundry. Third, the microphone must be local: it does not work over SSH, in cloud sessions, or in VS Code Remote sessions. In WSL it needs WSLg plus the sox and libsox-fmt-pulse packages.
Best for: anyone signed in with Claude.ai who runs Claude Code on the machine in front of them and is comfortable with cloud transcription. Nothing else on this page beats free and already installed.
3. Cloud-first apps: Wispr Flow and Aqua Voice
Wispr Flow runs on Mac, Windows, iOS and Android; its help center says there is no native Linux application. The free plan allows 2,000 words a week on desktop, and Pro costs $15 per user a month, or $12 a month billed annually. Flow needs a working connection: its troubleshooting steps start with restoring the internet connection, and Retry sends the saved audio again. In terminals it uses the same paste action as everywhere else and lists direct paste for Windows Terminal, PowerShell, Terminal.app, iTerm2, Warp, Ghostty, kitty and others. In WSL windows, Linux VMs and remote SSH sessions, Wispr says Flow still transcribes but you must use Paste last transcript. Its variable recognition reads names visible in Cursor, Windsurf and VS Code with Screen Reader Optimized mode on, so it does not apply to Claude Code in a plain terminal.
Aqua Voice is available for Mac, Windows and iOS, with Android on a waitlist. Free includes 1,000 words and the Avalon transcription model; Pro is $8 a month billed annually or $10 month to month, and Max, at $24 or $30, adds Realtime Mode and a “Send it” voice command. Aqua’s privacy policy describes transcription as a service that processes your audio and says transcripts may be stored on its servers unless Privacy Mode is on. Its homepage says it understands what is on your screen, from code syntax to everyday text.
Best for: people who want cloud accuracy, AI formatting and phone apps, and always work online. Neither runs on Linux.
4. Local-first apps: Superwhisper, VoiceInk, Spokenly and Handy
Superwhisper runs on Mac, Windows, iOS and Android with offline and cloud models. The free tier includes unlimited use of Whisper models; Pro, which adds cloud and local AI models without limits and your own API keys, is $8.49 a month, $84.99 a year or $249.99 once for a lifetime license, and you can try Pro for 3,000 words. Superwhisper says Intel Macs work best with cloud models. For Claude Code it offers a macOS add-on that notifies you when the agent finishes and lets you reply by voice, and a CLI for macOS and Windows that pipes transcripts into Claude Code.
VoiceInk is an open-source Mac app that transcribes locally and needs Apple silicon and macOS 14.4 or later. It is a one-time purchase: $25 for one Mac, $39 for two, $49 for three. Its optional Cloud Enhancement sends only transcribed text, never audio.
Spokenly runs on Mac, Windows, Linux and iPhone. Local Whisper and Parakeet models are free with no usage limits and no account, your own API keys are free too, and Pro with Spokenly’s cloud models is $99.99 a year. A local-only mode blocks every network request. The Linux version ships deb, rpm and AppImage packages for x86_64. On a Mac, Spokenly also runs an MCP server that Claude Code can call to ask you a question and take a spoken answer, with both apps on the same machine.
Handy is free and MIT-licensed, runs on Windows, macOS and Linux, and transcribes offline with Whisper or Parakeet V3. Its README says Wayland support is limited and that Linux needs a typing tool such as xdotool, wtype or dotool.
5. Operating-system dictation: free, but blind to code
macOS Dictation costs nothing and types into the focused text field, so try it in your terminal before buying anything. Apple says you can check Keyboard settings to see whether your general dictation is processed on your device rather than sent to Siri servers. Escape or the Dictation shortcut stops it, and it stops on its own after 30 seconds without speech.
On Windows, Win+H voice typing types wherever the cursor sits in a text box, but Microsoft says it uses online recognition powered by Azure Speech and needs an internet connection. On Linux, use one of the apps above that ships a Linux build.
Both are fine for an occasional sentence. Neither knows your repository, so expect file and function names to need retyping, and neither helps when Claude Code runs on another machine over SSH, beyond typing into the local window.
6. Where IraVoice fits, and where another pick is better
IraVoice 1.2.2 recognizes speech on your computer with NVIDIA Parakeet and runs on macOS 13+ on Apple silicon, Windows 10 or 11 (64-bit), and Linux (beta). It costs $79 once for one computer, $109 for two, or $149 for three, with no subscription, and the 3-day free trial starts with your first dictation. After the one-time model download it works offline, and audio and text never leave the computer.
It types into the focused terminal, so the same key works whether Claude Code runs locally, in tmux, in a WSL tab in Windows Terminal, or on a server you reached over SSH. It does not press Enter by default. On a Mac and on Windows it spells the code names of the Git project the terminal is working in, such as useEffect or Pipeline.swift, keeping only names and never source text. Over SSH, those names come from a project on your own computer, not the remote host; inside WSL they may not appear; and on Linux, a beta, there are none.
Choose something else if you need a phone app, an Intel Mac, a Linux build for ARM, or cloud recognition: IraVoice has none of those today.