Claude Code voice mode and privacy: where your audio goes

Where does Claude Code's voice mode send your audio?

To Anthropic. According to Anthropic’s documentation, as of 3 October 2026: “Voice dictation streams your recorded audio to Anthropic’s servers for transcription. Audio is not processed locally.”

How Anthropic keeps and uses that data is set out on its data usage page. Read it yourself before you decide: it is Anthropic’s policy to describe, and it can change after the date above.

You may be reading this because /voice is missing from your Claude Code, or because a contract or a policy makes you wary of uploading a recording of yourself reading out a client’s code.

Why is /voice not available for you?

Voice dictation needs a claude.ai account, and Anthropic’s page lists where it does not work:

• with an Anthropic API key configured directly;
• through Amazon Bedrock, Google Cloud’s Agent Platform or Microsoft Foundry;
• in a cloud session or an SSH session, since it needs a microphone on the machine running Claude Code;
• in the VS Code extension during a Remote session (SSH, Dev Containers, Codespaces), where the extension runs on the remote host and your microphone is on your desk;
• where an administrator has turned it off, with the message “Voice mode is disabled by your organization’s policy”.

A configured API key or third-party provider wins over a claude.ai sign-in, so /voice shows as an unknown command even while you are signed in. If you see that, look for a configured key or provider first.

Who might choose not to use it, even where it works?

People under an NDA or a company AI policy may hesitate. The upload adds less than it first sounds if the company approved Claude Code, because every prompt’s text reaches Anthropic when it is sent anyway.

The extra is the recording: your voice, plus anything you said and then deleted from the prompt, since the audio has already gone by the time you edit the text. Plenty of contracts say nothing about dictation, and a policy written with typed text in mind may not settle the question. If you cannot tell whether yours covers it, ask before you dictate a client’s material.

What does /voice do well?

If you have a claude.ai account, a local microphone and no objection to the upload, it is the easiest option, with nothing extra to install. Its transcription is tuned for coding vocabulary, so terms such as regex, OAuth, JSON and localhost are recognised, and it adds the current project name and git branch name as recognition hints automatically. Dictate Safe does not read your project or branch; you add those names to its vocabulary list yourself.

Transcription does not use Claude messages or tokens and does not count toward /usage limits. You can hold a key to record (Space by default) or tap to record and send; in tap mode recording stops after 15 seconds of silence or 2 minutes in total. It supports 20 dictation languages, English among them, which is 19 more than Dictate Safe.

Side by side: /voice and an on-device dictation tool

The middle column uses only Anthropic’s page. The last describes a tool that runs speech recognition on your computer and types into whatever field has focus.

What does dictating on your own computer change?

It moves one step onto your computer: turning your voice into text. The words you dictate are still sent to the model when you submit the prompt, as typed words would be, and your company’s rules for Claude Code still apply to them. If Claude Code itself is not allowed, no dictation tool changes that.

Dictate Safe, a dictation app for macOS and Windows, works this way: it runs its speech model on your own computer and types into whatever field has focus. It needs no account, works offline after a one-time speech model download, and is a single purchase. For how to phrase a spoken prompt, see how to dictate prompts to AI assistants or a day of prompting by voice.

What should you check before dictating client work?

Two checks fit this situation. Read where each tool sends audio, in its own documentation, and note the date: Anthropic’s voice page says it in 2 sentences. And ask whether your company’s approval covers the dictation tool as well as Claude Code. To test what a tool does, use a 10-minute offline test you can run on any app.

The post on whether dictation software sends your voice to the cloud explains the 3 kinds of tool and how to tell which one you have.