Voice for vibe coding.

Why voice and agents fit together

Andrej Karpathy coined "vibe coding" in February 2025, for the way of working where you describe what you want and let the model produce the code. Merriam-Webster listed it as slang and trending the following month, and Collins made it their word of the year for 2025.

The reason voice landed alongside it is mechanical, not fashionable. Typing code rewards precision: punctuation, symbols, exact identifiers. Describing intent rewards fluency, and speech is the faster channel for fluency. A prompt like "move the upload handler behind the queue, keep the existing error messages, and tell me what breaks in the integration tests" takes seconds to say and the better part of a minute to type. The research on how much faster is more equivocal than the marketing, and still lands on speech. Speaking it also tends to produce a better prompt, because you say the whole thought instead of the abbreviated version your fingers were willing to write.

The agents ship voice now

Claude Code has /voice, in hold-to-record or tap-to-record modes. Per Anthropic’s documentation, it "streams your recorded audio to Anthropic’s servers for transcription. Audio is not processed locally." Within its lines it is well built: 20 languages, transcription tuned for coding vocabulary, project and branch names fed in as hints, and no token cost. The lines themselves are narrow, and rather than repeat them here, the /voice comparison lists each one against its source.

Cursor built voice in too. Their 2.0 changelog describes it as "Control Agent with your voice using built-in speech-to-text conversion", with custom submit keywords you can define. Version 3.1 sharpened it: hold Ctrl+M to speak, a waveform and timer while you record, and the full clip transcribed "with batch STT for higher-quality speech-to-text". Their changelogs frame it around the Agent and the Agents Window, and say nothing about where the audio is processed.

These are good features and worth using. If your setup fits inside their boundaries, you don’t need anything else, and that is the honest recommendation.

Where the built-ins stop

Each built-in serves its own prompt box. That’s a sensible scope, and it’s also the boundary you hit first. A day of agent work isn’t only the agent’s input field: it’s the commit message, the pull request description, the Slack reply explaining what you just shipped, the terminal in the next tab running a different tool, the follow-up question the agent asks mid-task. A system-wide dictation app covers all of it with one key, because it types wherever the cursor already is.

The other boundaries are structural. Anthropic requires a Claude.ai account for /voice, so the feature disappears the moment you authenticate another way; a dictation app never asks, because text typed into a terminal is only text. SSH is the same story from a different angle. Anthropic’s docs rule out remote environments, while a local app types into your local terminal and the remote session receives ordinary input, so it doesn’t know the difference. Then there’s the audio itself, which Anthropic says plainly is not processed locally. Fine for most work. Disqualifying for some.

What on-device changes

Edict is the version of this that never sends audio anywhere: the speech models ship inside the app, so it transcribes on your Mac and works offline from the first launch. For agent work that matters in specific ways. It’s verbatim, applying only your own dictionary, so the prompt reaches the agent as you said it rather than as some model paraphrased it. There’s a wrong-window guard, so a forty-second prompt doesn’t land in Slack because focus moved while you were talking. And the MCP server turns the arrangement around: instead of you driving dictation, the agent can read your transcript history, edit your dictionary, or transcribe a file locally on your behalf.

The full walkthrough for Claude Code covers setup, auto-send, and SSH, and the side-by-side with /voice has the comparison table. Edict’s half of both carries over to Cursor, Codex, or anything else with a text input, since typed text is the one interface they all share. Anthropic’s constraints are Anthropic’s alone, and every agent’s built-in comes with its own.

Common questions