Talk to Claude Code instead of typing
By Dmytro Lakiza, who makes Shepit, so one of the two options below is his. Claude Code details are from Anthropic's voice dictation docs, read on 18 November 2026.
I'm a product designer and I build products with a small team of Claude Code agents. The biggest change in how I work wasn't a model. It was that I stopped typing prompts. A spoken prompt comes out much longer than a typed one, and the extra context is usually what the agent needed: the reason behind the change, the edge case I'd have skipped, the thing I tried yesterday that didn't work.
On a Mac there are two good ways to do it.
Option 1: Claude Code's built-in /voice
Run /voice in the CLI, then hold Space (or tap it) and talk. Your words appear in the prompt live, and you can mix voice and typing in the same message. It's tuned for coding words like regex, OAuth and JSON, and it adds your project and branch names as hints. Transcription doesn't use your Claude messages or tokens. It also works in the VS Code extension.
Things to know, from Anthropic's docs:
- The audio streams to Anthropic's servers for transcription; it isn't processed locally.
- It needs a Claude.ai login. It isn't available if you use an API key, Bedrock, Google Cloud or Microsoft Foundry.
- It doesn't work in SSH, cloud or VS Code Remote sessions.
- It uses one dictation language, the same one Claude replies in, and falls back to English if yours isn't supported.
- It works inside Claude Code only.
If you live in one terminal with a Claude.ai account and speak English, start here. It's already installed.
Option 2: dictation for the whole Mac
A system-wide dictation app puts text wherever your cursor is: the Claude Code terminal, Codex, Cursor, a GitHub comment, a Slack reply to the person who asked why the build broke. You hold one key everywhere and stop thinking about which tool has which mic button.
That's what I built Shepit for. Speech recognition runs on the Mac, so the audio never leaves it and it works offline once the models are downloaded. On an M3 Pro, 27 seconds of speech becomes text in about 0.3 seconds. It detects 25 European languages as you speak (99 with Whisper models), and it copes with English terms dropped into the middle of another language. That's how I talk to agents: Ukrainian sentences full of English words like "refactor", "endpoint" and "commit". We've measured that mix on Ukrainian only, so try it on yours.
Side by side
| Shepit | Claude Code /voice | |
|---|---|---|
| Where it works | Any Mac app | Claude Code CLI and the VS Code extension |
| Speech recognition | On your Mac | Anthropic's servers |
| Language | Detected as you speak (25 with Parakeet), 99 with Whisper | One configured language, English by default |
| Account | None | Claude.ai login |
| Also does | Notes from your calls with who said what, made on the Mac | Coding vocabulary and project-name hints |
| Price | Free to start: 5,000 dictated words a week. Pro $7 a month, $59 a year, or $79 once (first 200 buyers) | Included with Claude Code, doesn't count toward usage |
Tips that work with either
- Say the goal first, then how the agent can check it's done. The check is the part people skip when typing.
- Don't clean up as you go. Ramble, then let the model sort it out; it's better at ignoring tangents than you are at guessing which detail matters.
- Read it before you send. Spoken prompts sometimes change their mind halfway through.
Try it
Shepit is free to start, and every install begins with 14 days of Pro (pricing). Hold the key in your terminal and describe the next change out loud.
Get Shepit for MacSources
Anthropic: Claude Code voice dictation docs (read 18 Nov 2026). Shepit: shepit.lakiza.io, privacy policy. Ukrainian accuracy: models.handy.computer.