TL;DR Enable
/voice, hold the spacebar and speak. Claude Code transcribes in streaming — you see words appearing as you talk. Release and the prompt is ready. Mix voice and typing in the same message.
There's a clear trend in the AI agent world: voice interaction. Typing long, descriptive prompts in a terminal is slow. Dictating them is more natural, faster, and keeps your hands free to browse code or documentation in another window.
Claude Code has incorporated this with streaming transcription — you don't wait until you finish speaking to see text. Words appear in real time as you dictate, optimized for development vocabulary: regex, OAuth, JSON, localhost are transcribed correctly. It even picks up your project name and current git branch as recognition hints.
In my experience it still works a bit rough around the edges — sometimes the "hold to record" detection lags, or the transcription doesn't nail very specific terms. But it's good to have as an option, and it will improve.
Result:
> /voice
Voice mode enabled. Hold Space to record. Dictation language: en
> [hold Space, speak: "refactor the auth middleware
to use the new token validation helper"]
> refactor the auth middleware to use the new
token validation helper▮
Setup
1. Enable Voice Mode
/voice
Persists across sessions. Run /voice again to disable.
2. Speak
Hold Space and talk. You'll see keep holding… briefly, then a live waveform. Release to finalize.
Text inserts at the cursor position — you can dictate at any point in the prompt and combine with typed text.
3. Change the dictation language
Dictation uses your language setting. If not defined, defaults to English:
{
"language": "es"
}
4. Rebind the push-to-talk key (recommended)
Space has a warmup because it needs to distinguish between holding and typing. A modifier combination starts instantly:
// ~/.claude/keybindings.json
{
"bindings": [
{
"context": "Chat",
"bindings": {
"meta+k": "voice:pushToTalk",
"space": null
}
}
]
}
Reference
| Aspect | Detail |
|---|---|
| Enable | /voice (toggle) |
| Default key | Space (hold) |
| Transcription | Streaming in real time |
| Vocabulary | Optimized for development (regex, OAuth, JSON, etc.) |
| Auto hints | Project name + current git branch |
| Languages | 20 languages (en, es, fr, de, ja, ko, pt, etc.) |
| Persistence | Stays enabled across sessions |
| Requirements | claude.ai login (not API key), local microphone |
| Minimum version | v2.1.69+ |
Related: dictation is the tool for a bigger method — reading a huge change out loud, in concentric circles, until you understand it.
Official docs: Voice dictation
Requirements
- Claude Code v2.1.69+
- claude.ai account (doesn't work with API key, Bedrock, Vertex, or Foundry)
- Local microphone access (doesn't work over SSH or remote environments)
- macOS/Linux/Windows (Linux may need SoX or ALSA utils)
At the opposite end, if your thing is never leaving the keyboard, Claude Code ships vim mode inside the prompt.
