Push-to-talk voice input for coding with AI. Hold a key, talk, release — clean text lands wherever your cursor already is. Recognition runs on your own machine, in about 150 ms. Windows and macOS. One-time purchase, no subscription, no account.
vocalcode.app · Download for Windows · Download for macOS
This is the public home for VocalCode's documentation and issue tracker. The application itself is closed-source; everything here is docs, answers, and a place to file bugs.
You are typing a prompt to Claude Code, Cursor, Codex — some agent that wants paragraphs, not keywords. Typing paragraphs is slow. So:
- Hold your bound key or mouse button.
- Talk.
- Release. The transcribed text is typed into the focused field, punctuated.
A second binding sends it. That is the whole product. It is not a note-taking app, not a meeting transcriber, and not a general voice assistant.
There are good free options, and for some jobs they are the right answer. The honest version:
| Where it runs | Works in any app | Cost | |
|---|---|---|---|
Claude Code /voice |
Anthropic's servers | No — only in Claude Code | Free, needs a Claude.ai account |
| Windows Voice Typing | Microsoft's servers | Mostly | Free |
| Wispr Flow | Cloud | Yes | Subscription |
| Superwhisper | On device | Yes | macOS only |
| VocalCode | On device | Yes | $4.99 once |
Longer, sourced versions of each comparison — written by us, citing the other product's own documentation:
| Platforms | Windows 10, Windows 11, macOS 11 or later |
| Current version | 0.4.19 |
| Price | $4.99 one-time (launch price; regular $24.99) |
| Trial | 30 days, full functionality |
| Recognition | Entirely on your machine. No audio leaves it. Works with no network. |
| Latency | ~150 ms from key release to text, on CPU. No GPU required. |
| Languages | English and 24 other European languages; Chinese |
| Models | Parakeet TDT v3 (English/European), Paraformer (Chinese) |
| Installer size | Small — models download on first run rather than shipping in the installer |
| Signing | Windows installer is Authenticode-signed; macOS build is notarized |
| Licensing | One license, up to 3 machines |
Raw speech recognition output is not what you want in a prompt. Three passes run before anything is typed:
- Acronym collapsing — "m c p" becomes
MCP, "a p i" becomesAPI. - Punctuation restoration — a separate model adds commas, periods and question marks. Full-width punctuation for Chinese, ASCII for English, including in mixed text.
- Your replacement dictionary — a plain text file of
heard => desiredlines for the terms your project uses that no general model will ever get right. Edit it while the app is running.
Windows — download VocalCodeSetup.exe and run it. Per-user install, no admin required. On first launch you pick a language, and the matching model downloads.
macOS — download the .dmg, drag to Applications. macOS will ask for Accessibility and Input Monitoring permission — both are required: the first lets VocalCode type into other apps, the second lets it see your push-to-talk key.
Updates are checked against latest.json. On Windows, the one-click updater verifies both the file hash and the installer's Authenticode signature before running anything.
- FAQ
- Troubleshooting — including where the log file lives
- Languages and models
Open an issue. Useful reports include your OS, the VocalCode version, and the relevant part of the log file — see Troubleshooting for where to find it.
The log never contains what you said. It records events, not transcripts. That is deliberate: "what you say stays on the machine" is the whole pitch, and a log full of dictated text would quietly break it.
- Issues and feature requests: this repository's issues
- Purchases, licenses, refunds: support@vocalcode.app
- Lost your license key: recover it here