🎙 Voice Prompt for Claude Code

Talk to your terminal. Claude Code listens.

claude-voice.mov

Especially useful if you use Claude Code from the terminal directly. Records your voice, transcribes it locally with Whisper (OpenAI's offline speech model — audio never leaves your Mac), optimizes the text automatically, and pastes it into Claude Code with a single keypress. If you're on Ghostty, /voice opens a split panel automatically alongside your chat.

Made with ♥ by Guille Varela · Nodox Studio

---

Demo

Type /voice inside Claude Code — a split panel opens right alongside your chat:

┌──────────────────────────────────────────────────────┐
│  🎙  Voice Prompt  ·  Claude Code                    │
│  ────────────────────────────────────────────────    │
│  by Nodox Studio  ·  Ctrl+C to exit  ·  Cmd+W close │
│                                                      │
│  ►  Press Enter to record…                          │
│                                                      │
│  🔴 Recording…  press Enter to stop                 │
│                                                      │
│  ⟳  Transcribing…  [Whisper — local]               │
│                                                      │
│  🎤 Dictated:    uh can you fix the bug in the      │
│                  login form, like, you know          │
│  ✨ Optimized:   Fix the bug in the login form.     │
│                                                      │
│  1 🚀 Send to Claude Code                           │
│  2 ✨ Copy optimized prompt                         │
│  3 ✏️  Edit text                                     │
│  4 🔴 Record again                                  │
│                                                      │
│  Press a number…  1                                 │
│                                                      │
│  ✓  Sent to Claude Code.                            │
└──────────────────────────────────────────────────────┘

---

What the optimizer does

No AI — just fast, deterministic local rules:

You say	Claude receives
"uh can you fix the bug in the login form"	Fix the bug in the login form.
"could you please refactor the nav component"	Refactor the nav component.
"can you explain me how the auth system works"	Explain how the auth system works.
"i need you to create an endpoint, and also add tests, and also document it"	1. Create an endpoint. / 2. Add tests. / 3. Document it.

Rules applied on-device, no API call, no latency:

1. Removes filler words — uh, um, like, you know, eh, o sea, bueno, pues…
2. Strips polite preambles — can you, could you, quiero que, necesito que…
3. Normalizes Spanish verbs — que refactorices → Refactoriza, explicarme → Explica
4. Multi-part → numbered list — detects and also, y también, y además…
5. Capitalizes and adds end punctuation

---

Requirements

• macOS 13+
• Python 3.10+
• Homebrew
• Claude Code CLI
• Ghostty terminal (for the /voice split panel — other terminals work fine, just without the auto-open)

---

Installation

git clone https://github.com/nodox-studio/claude-voice
cd claude-voice
chmod +x install.sh
./install.sh

The installer walks you through everything: what gets installed, what permissions it needs, and what stays on your machine. It asks before doing anything.

macOS permissions

You'll need to grant two permissions:

• Microphone — macOS will show a dialog on first recording
• Accessibility — needed for auto-paste into Claude Code → System Settings → Privacy & Security → Accessibility → enable your terminal app

---

Usage

From inside Claude Code (recommended):

Type /voice — opens a vertical split panel in Ghostty with the tool running right alongside your chat.

From any terminal:

voice

Workflow:

1. Press Enter → starts recording
2. Speak your prompt naturally (filler words, incomplete sentences — doesn't matter)
3. Press Enter → stops, transcribes, and optimizes
4. Pick an action with a single keypress — no Enter needed

Press Ctrl+C to exit · Cmd+W to close the split panel

---

Privacy

This was a design priority, not an afterthought.

✅	Audio never leaves your Mac — Whisper runs entirely on-device
✅	Temp files deleted immediately after every transcription, even on failure
✅	No telemetry, no analytics, no accounts — it's just a script
⚠️	Google fallback — if Whisper fails with a technical error, the app falls back to Google Speech API and tells you explicitly on screen. You can remove this fallback in voice_prompt.py if you want hard-fail instead.

---

Configuration

Edit these constants near the top of voice_prompt.py:

WHISPER_MODEL = "tiny"     # tiny (~75 MB) · base (~150 MB) · small (~500 MB)
TERMINAL_APP  = "Ghostty"  # Ghostty · iTerm2 · WezTerm · Terminal

tiny is fast and works well for short prompts. Upgrade to base or small if you need better accuracy with heavy accents or complex vocabulary.

TERMINAL_APP controls which terminal receives the auto-paste. The /voice split panel is Ghostty-only, but recording and paste work with any terminal listed above.

---

Security

• All subprocess calls use argument lists — no shell=True, no injection risk
• Transcribed text is never passed as a shell argument — it goes through pbcopy (stdin) and is pasted with a simulated Cmd+V
• Temp audio files are deleted in a finally block, even if transcription fails
• No credentials, API keys, or tokens required or stored
• Auto-paste uses AppleScript keystroke simulation, which requires Accessibility permission granted explicitly by you in System Settings

---

Contributing

Found a bug? Have an idea? PRs are very welcome.

• 🌍 More languages — Whisper supports 99 languages. Add your language's filler words to the optimizer in voice_prompt.py
• 🖥 Other terminals — iTerm2, WezTerm, kitty… the main thing to swap is the AppleScript split navigation
• 🐛 Issues → open a GitHub issue or email somos@nodox.studio

---

Support

If this saves you time and you feel like it — no pressure at all:

☕ Buy Guille a coffee

---

License

MIT © Guille Varela / Nodox Studio