Claude Code has dictation now. When do you need more?

Claude Code ships voice dictation. Run /voice, hold Space, talk, release. It is tuned for coding vocabulary — regex, OAuth, JSON, localhost — and it quietly adds your project name and git branch as recognition hints. It costs nothing and consumes no tokens.

If you are here deciding whether you still need a separate dictation app, here is the short answer: for dictating prompts to Claude Code, often you do not. It is a good implementation of exactly that job, it is free, and it is already installed.

The rest of this is the two places where that stops being true — and both of them are stated in Anthropic’s own documentation rather than inferred by us.

What the built-in does, accurately

Worth stating properly, because a fair comparison starts with the other thing being good.

/voice toggles it on. Hold mode is the default and is real push-to-talk: recording runs while you hold Space and stops when you release. /voice tap switches to tap-to-start and tap-to-send, which skips the brief key-repeat warmup that hold mode needs. You can rebind the key to something like meta+k. Your speech appears in the prompt dimmed until it finalises, inserts at the cursor, and you can mix typing and dictation in the same message. Twenty languages. Recording stops on its own after fifteen seconds of silence or two minutes total.

That is a well-built feature. If your dictation begins and ends inside Claude Code’s prompt, the honest recommendation is to use it.

The first fact: the audio leaves your machine

From the Claude Code documentation, quoted exactly (retrieved 2026-09-11):

“Voice dictation streams your recorded audio to Anthropic’s servers for transcription. Audio is not processed locally.”

That is a clear disclosure and it deserves credit for being unambiguous — a great many products are vaguer about this. It is also, for a particular group of people, the end of the conversation.

If you work under a contract that forbids third-party processing of audio from the machine, or you dictate anything under NDA, or you simply hold the position that speech from your desk is not something you rent out, then no amount of quality in the feature changes the answer. The constraint is not “is it good”, it is “may the audio leave”.

Halopen’s local log stores transcripts and supported app events. A local dictation entry identifies its transcription route; it is not a monitor of every app or Mac network request. Account, licensing, updates, usage and enabled diagnostics have separate network behavior. See the privacy page for processing and retention details.

The second fact: it works in one place

The built-in dictates into Claude Code’s prompt input, and into the VS Code extension. The docs also rule out remote environments — SSH sessions and Claude Code on the web — because the microphone is on your machine and the session is not.

So the question is not really “which dictation is better.” It is how many dictation habits do you want to have.

A normal hour is not spent in one window. It is a prompt to an agent, then a commit message, then a reply in Slack that has to sound like a person, then a comment on a pull request, then a note to yourself you will not otherwise remember. If dictation only works in one of those, you type the other four — and a habit that works four-fifths of the time does not become automatic. That is the actual cost, and it is easy to underrate until you have felt the reach-for-a-key reflex arrive in the wrong app and die.

Halopen holds the same key everywhere a cursor goes. It also knows where it is: in Terminal, iTerm2, Warp, Ghostty, VS Code, Cursor, Zed and Xcode it drops any cleanup mode to verbatim automatically, because a shell command is not a sentence that benefits from being improved.

A third, narrower one worth knowing

The docs state that the speech-to-text service requires a Claude.ai account and “is not available when Claude Code is configured to use an Anthropic API key directly, Amazon Bedrock, Google Cloud’s Agent Platform, or Microsoft Foundry.”

If you are at a company, there is a decent chance you are on one of those paths and the /voice command is simply not available to you. An administrator policy can also turn it off for an organisation. Worth checking before you plan around it.

So: a straight recommendation

Use the built-in if your dictation is mostly prompts to Claude Code, you are signed in with a Claude.ai account, and sending audio to a transcription service is fine for your work. It is free and it is good. We would rather you used it than bought something you do not need.

Look further if any of these is true:

1. The audio cannot leave the machine. This is the one that admits no workaround.
2. You dictate in more than one app. One habit that works everywhere beats a better habit that works in one window.
3. You authenticate Claude Code with an API key, Bedrock, Vertex or Foundry, and /voice is therefore unavailable to you.
4. You want the words you actually said. A prompt is context: a hedge is a constraint and an intensifier is a severity signal, and tidying them away changes what the sentence claims.

That is the whole decision, and for a lot of readers it lands on “the built-in is fine.” Anthropic shipped a good free feature and the honest thing is to say so. Recommending against ourselves in point one of four costs a sale we were not going to deserve, and it is the only version of this page worth publishing.

---

Every statement about Claude Code above is quoted from or directly supported by Anthropic’s voice dictation documentation, retrieved 2026-09-11. Halopen is native macOS dictation, verbatim by default, with an on-device mode that recognizes speech locally, with separate account, update and diagnostic networking. Try the time-limited no-card trial — halopen.com.

Current trial duration and paid plans