Features
Pricing

Claude Code voice, VS Code Speech, Cursor or a dictation app?

A dated table, from each vendor's own documentation, of where Claude Code, VS Code, Cursor and Dictate Safe, a dictation app, process your voice and what each needs.

The Dictate Safe team8 min read

In short

  • Anthropic's documentation, checked on 3 October 2026, says Claude Code's /voice streams your audio to Anthropic's servers, needs a claude.ai sign-in and does not work with an API key or over SSH.
  • On the desktop, VS Code's experimental built-in dictation transcribes on your computer by default, and its default cleanup setting sends the transcript text to a Copilot language model.
  • The VS Code Speech extension, which VS Code recommends for Intel Macs, says audio is processed locally and no internet connection is required.
  • Cursor documents a microphone icon in the chat input, and none of its pages we read says where that audio is processed.
  • Dictate Safe, a dictation app for macOS and Windows, transcribes on your computer and types into any program with a text field, in English only.

What is being compared, and how fresh is this?

This compares five ways to dictate a prompt into an AI coding tool, as each vendor’s documentation described them on 3 October 2026: Claude Code’s /voice, VS Code’s built-in dictation, Microsoft’s VS Code Speech extension, Cursor’s voice input, and Dictate Safe, a dictation app that types into any program. We did not run the other vendors’ tools, and nothing here rates how well any of the five recognises speech.

How do the five options compare?

ToolWhere it worksWhere speech becomes textWith the network off
Claude Code /voiceThe CLI on macOS, Linux and Windows, and the VS Code extension. Not over SSH, in cloud sessions or in VS Code Remote.Anthropic’s servers: “Audio is not processed locally.”Needs a connection.
VS Code dictationChat, the Agents window, editors and terminals, including remote workspaces. On-device on Windows, Apple silicon Macs and Linux with glibc 2.34 or later.Desktop: your computer, by default. Web: the Microsoft AI voice service. A default-on cleanup setting sends the transcript text to a Copilot language model.Desktop: works after the first model download. Web: needs a connection.
VS Code Speech extensionChat (with GitHub Copilot Chat) and the editor, on Windows, macOS and listed Linux distributions. VS Code sends Intel Macs, 32-bit and Arm32 systems and musl Linux here.“Processed locally on your computer.”“No internet connection is required.”
Cursor voice inputThe chat input, and Design Mode in the Agents Window.Not stated in the documentation we read.Not stated in the documentation we read.
Dictate SafeAny program with a text field, on macOS 12 or newer and Windows 10 (64-bit) or newer.Your computer, with an on-device speech model. A cloud option using your own key is off by default.Works after a one-time speech model download (about 150 MB for the default).
ToolWhat you needCost, as documentedHow you start it
Claude Code /voiceA claude.ai sign-in on Pro, Max, Team or Enterprise, and a local microphone. Not with an API key, Amazon Bedrock, Google Cloud’s Agent Platform or Microsoft Foundry.Transcription uses no messages or tokens and does not count toward /usage limits./voice, then hold Space; or /voice tap, then tap Space to start and again to send.
VS Code dictationAI features turned on. VS Code for the Web also needs a GitHub sign-in.Not stated in the documentation we read.Chat: the microphone button or Cmd+I (Ctrl+I). Editor: Option+Cmd+V (Ctrl+Alt+V). Terminal: a Command Palette command.
VS Code Speech extensionThe extension. Chat also needs GitHub Copilot Chat and a sign-in.No charge, according to its Marketplace listing.Press or hold Cmd+I (Ctrl+I) in chat, or Cmd+Alt+V (Ctrl+Alt+V) in the editor. VS Code’s page adds a “Hey Code” keyword.
Cursor voice inputNot stated in the documentation we read.Not stated in the documentation we read.“Click the microphone icon in the chat input”.
Dictate SafeThe app and one speech model. No account is needed to dictate.One payment, no subscription (see pricing).Hold the dictation key (Fn by default on a Mac, F9 on Windows) or click the floating icon.
ToolLanguagesLimits the documentation states
Claude Code /voice20, with English the default.In tap mode, recording stops after 15 seconds of silence or 2 minutes. An administrator can turn voice off.
VS Code dictationNo list. A setting gives a language hint.Experimental, and it “might change or be removed”. Stops after 20 minutes.
VS Code Speech extension26.“This extension is still in development”.
Cursor voice inputNot stated in the documentation we read.Not stated in the documentation we read.
Dictate SafeEnglish only.No Linux. Project and branch names go in its vocabulary list by hand.

Checked on 3 October 2026. Sources: Anthropic’s voice dictation and setup pages; Microsoft’s VS Code voice support page and the VS Code Speech Marketplace listing; Cursor’s pages on prompting agents and Design Mode. The Dictate Safe rows are the maker describing its own product.

When is the built-in option the right one?

Often: if you dictate into one tool and its documented behaviour suits your work, use what is already there.

  • Claude Code is the only place you dictate, your plan includes it, and your contracts allow the recording to go to Anthropic. /voice needs no install and adds your project and git branch names as recognition hints. We would not suggest buying anything for that.
  • You use VS Code on the desktop and want the audio kept local. Its built-in dictation runs on the device by default, remote workspaces included; on an Apple silicon Mac we would try it first.
  • Your Mac has an Intel processor: VS Code sends you to the Speech extension, whose listing says audio is processed locally.
  • You dictate in a language other than English: Claude Code lists 20 languages and the Speech extension 26.

If Cursor’s microphone button does the job, keep using it. Before you dictate a client’s material into it, ask Cursor where the audio goes or run the 10-minute offline test.

When does an app that works in any program make sense?

When you dictate into more than one program, or the audio must stay on your computer in every one of them. Dictate Safe, a dictation app for macOS and Windows, runs speech recognition on your computer and types into whichever text field has focus, in VS Code, Cursor, a browser tab or Slack. It also fits if:

  • Your Claude Code uses an API key, Amazon Bedrock, Google Cloud’s Agent Platform or Microsoft Foundry, so /voice is unavailable.
  • You want one switch for the network: Offline Lock refuses every outbound connection the app makes, ours included.
  • You dictate code: Code Mode, off by default, types ( when you say “open paren”.

On a Mac, a setting that is on by default lets Dictate Safe find the focused text field in Electron apps such as VS Code and Cursor (see the help article). We have not tested their integrated terminals or the Claude Code extension’s panel, so try yours first.

It is the wrong pick on Linux or for a language other than English. For a spoken conversation with an agent, VS Code’s experimental Voice Mode reads replies aloud; it needs an eligible individual GitHub Copilot plan.

How do you decide in two minutes?

Answer these in order.

  1. Must the audio stay on your computer? Then rule out /voice and VS Code for the Web, and hold off on Cursor’s button until Cursor says where its audio goes.
  2. Do you dictate into one tool or several? For one, use its built-in option if it passed the first question. For several, an app that types anywhere saves learning a different button in each.

If two options survive, give the one already installed a day of real prompts before you buy anything, ours included.

What could we not check?

Cursor’s documentation is the biggest gap. Its editor pages mention the chat microphone and narration in Design Mode, and our search for “voice”, “dictation” and “microphone” found none that says where the audio is processed, whether it works offline, which languages it handles or what it costs. Microsoft’s page does not say whether VS Code’s desktop dictation needs a GitHub account or what it costs.

When we read them, VS Code’s page was dated 30 September 2026 and the Speech listing showed version 0.16.0, dated 13 June 2025 in its version history. Where a vendor’s page and this table differ, trust the vendor, and tell us through the contact page.

Questions people ask

Is Claude Code's /voice processed on my computer?

No. Anthropic's documentation, checked on 3 October 2026, says Claude Code's voice dictation streams your recorded audio to Anthropic's servers and that audio is not processed locally. It needs a claude.ai account and a microphone on the machine running Claude Code.

Does VS Code dictation work offline?

On a supported desktop, yes, according to Microsoft's VS Code documentation checked on 3 October 2026: the experimental built-in dictation downloads a speech model on first use, then works offline. VS Code for the Web streams audio to the Microsoft AI voice service, and a cleanup setting, on by default, sends the transcript text to a Copilot language model.

Does Cursor's voice input send audio to the cloud?

Cursor's documentation, as we read it on 3 October 2026, says to click the microphone icon in the chat input to dictate a prompt, and none of its pages we read says where the speech becomes text. Ask Cursor, or switch the network off and see whether it still transcribes.

Does Dictate Safe work in VS Code and Cursor?

Dictate Safe, a dictation app, types into whichever text field has focus. On a Mac, VS Code and Cursor are Electron apps, and the setting that lets it find their focused field is on by default; Windows needs no setting. We have not tested their terminals, so try it in the panes you use before relying on it.

Keep reading