How to Dictate in Cursor: Every Input, Not Just the Agent

TL;DR: To dictate in Cursor, run a system-wide speech-to-text app, put your cursor in any Cursor input — the Agent or Composer prompt, a Cmd+K inline edit, the chat panel, the integrated terminal, or the code editor itself — hold your dictation hotkey, and speak. Cursor has its own voice mode (start and stop it with Cmd+Shift+Space, then send with Cmd+Return), but it drops text into the Agent prompt and nowhere else. A system-wide voice-to-text tool like Voibe types into every surface, can keep your code on your machine, and resolves your project’s file and folder names as you speak. This guide covers both paths on Mac, Windows and Linux.

This is the killer use case for voice on a Mac: dictation is faster than typing for prose, and AI prompts are prose. Speaking a multi-file Composer instruction is far quicker than typing it — once your tool can spell your file names correctly.

Cursor is an AI-first code editor (a fork of VS Code) with several places you type — and therefore several places you can dictate. A system-wide dictation tool works in all of them because it inserts text wherever your cursor is, exactly like a keystroke:

Cursor's Cmd+K turns a selection into a diff from a natural-language instruction, Composer (Composer 2 shipped in 2026) makes coordinated edits across multiple files, and Agent can read the codebase, run shell commands, and iterate until a task is done. All three are driven by prose you write — which is exactly what dictation accelerates. The catch is that Cursor's built-in voice mode only fills the Agent prompt, so to dictate into Cmd+K, the chat, or the editor itself you need a system-wide tool.

Cursor shipped a native voice mode in 2.0, and it is still here in the 3.x builds. You start it, speak, stop it, and the text lands in the Agent input. It is genuinely useful for firing off a quick agent prompt without typing. But it is scoped narrowly, and three things about it matter before you lean on it.

It only targets the Agent prompt. Cursor’s voice mode does not type into Cmd+K, the chat panel, the editor, or the terminal, and users on the Cursor community forum report it does not support @-file mentions, model selection, or mode switching by voice. It is prompt dictation, not editor control.

There is no live transcript. While you are speaking you get a microphone waveform and nothing else — the transcription only appears once you stop the mic. If you are used to watching words land as you talk, the silence reads like a failure when it is just how the feature works.

Its processing location is undocumented. Cursor has not published whether its voice transcription runs on-device or streams audio to a server, and there is no voice section in Cursor’s docs to check. For a tool you point at a proprietary codebase, that matters: your spoken prompts routinely contain file names, function names, and architecture details. A cloud transcription step sends all of that off your machine.

A system-wide tool addresses all three. It types into every surface (and every other app), it shows you the text where you are typing, and it gives you a choice of where audio is processed — on-device, where nothing leaves your machine, or a private zero-retention cloud that never stores or trains on your data. Here is the honest comparison:

The pragmatic answer: use Cursor’s native voice for throwaway agent prompts if you like it, and a system-wide tool as your primary driver for everything else. Here is how to set up the system-wide path.

This trips up more people than any other part of Cursor voice input, because the control you would reach for — saying “submit” — is the one that does not work. Here is the actual sequence.

The submit keywords do not work. Cursor lets you configure a spoken keyword such as “submit” in settings, and it should end the recording and send. As reported on the Cursor community forum against version 3.11.19, saying it “does nothing; the mic just keeps listening,” even after several seconds of silence. The reporter notes this is a regression of an earlier issue that had been fixed in 2.0.32. Until it is fixed again, the working sequence is the one above: stop with Cmd+Shift+Space, then send with Cmd+Return.

None of this applies to a system-wide dictation tool, which is worth saying plainly because the two get conflated. There, the hotkey you hold is the recording, releasing it inserts the text at your cursor, and sending is whatever that app’s send key already was — there is no separate voice state to get stuck in.

Any system-wide Mac dictation tool will type into Cursor. This guide uses Voibe because it can run entirely on-device on an Apple Silicon Mac and has the workspace-aware Developer Mode that makes voice-prompting Cursor reliable; the same steps apply in spirit to alternatives covered at the end.

1. Download Voibe from getvoibe.com (or the direct .dmg) and drag it to Applications.
2. Launch it. On Apple Silicon (M1–M4, macOS 13+) it downloads a local Whisper model (~2 GB) on first run.
3. There is no account to create and no internet needed after the model downloads.

For the full install walkthrough including the first-launch security prompt, see our Voibe setup guide.

A system-wide tool needs two macOS permissions to work inside Cursor:

1. Accessibility — lets the tool insert text into Cursor (System Settings > Privacy & Security > Accessibility, then enable the app).
2. Microphone — lets it capture your speech (granted on first use, or System Settings > Privacy & Security > Microphone).

Then pick a hotkey you can comfortably hold while your hands rest on the keyboard. This is push-to-talk: the microphone is live only while the key is down, so there is no voice state to toggle in and out of and nothing listening when you are not holding it. Voibe defaults to holding the Fn key: press and hold, speak, release, and the text appears at your cursor. If Fn conflicts with your keyboard layout, set a different key in settings — a held right-Command or a custom combination works well. For a deeper look at hotkey options and conflicts, see our Mac dictation keyboard shortcuts guide.

Hold-to-talk is the right pattern for coding: you keep one hand on the hotkey, dictate a prompt or comment, release, and your hands are already back on the keyboard to edit.

This is the step that makes voice-prompting Cursor actually work, and it is the feature no native editor voice mode offers. Open Voibe's settings and toggle Developer Mode on. Once enabled, Voibe detects your open Cursor (or VS Code) window and resolves file names, folder names, and project-specific terms from your workspace as you dictate.

The problem it solves: spoken identifiers don't transcribe cleanly. Say "update user service dot t s" and a normal dictation tool writes exactly that. Developer Mode matches the words against your actual workspace and inserts userService.ts. The same applies to folders, components, and any term that appears in your project tree.

Why it matters for Cursor specifically: the single biggest friction in voice-prompting an AI editor is getting it to reference the right files. A prompt like "refactor the auth handler in userService and update its test" only works if "userService" lands as the real filename. Developer Mode turns that from a manual cleanup step into something automatic.

Here is what dictating into each Cursor surface looks like in practice. Hold your hotkey, speak the instruction, release, and (where relevant) submit.

Cmd+K — Targeted Inline Edit

Select a block of code, press Cmd+K, then dictate the change:

• "Refactor this function to use async/await instead of callbacks, and add error handling with a try-catch block."
• "Extract these three validation checks into a helper called validateInput and call it here."

Composer — Coordinated Multi-File Change

Open Composer and dictate a change that spans files. This is where voice wins most, because multi-file instructions are long to type:

• "In the checkout flow, add a loading state to the Pay button and disable it while the request is in flight. Update the button component and the checkout page that uses it."

Agent — Delegated Task

Dictate a task and let the agent plan and execute:

• "Add a rate limiter to the public API routes, write tests for it, and run them."

The Editor and Terminal

Outside the AI surfaces, dictate directly: a docstring above a function, a Markdown section in your README, or a commit message in the terminal ("fix: prevent duplicate submissions on the checkout button").

Keep prompts to short, structured chunks — voice transcription handles a focused two-sentence instruction far better than a 60-word run-on. For a repeatable structure to dictate by, use the Five-Part Voice Prompt framework (Goal, Inputs, Constraints, Example, Output) from our voice-prompt AI guide, and pair it with the Talk-Draft-Polish loop in our voice input workflow guide.

If the same voice workflow also drives Claude Code, there is one prompt worth recognising before you answer it on autopilot: the post-rating question asking whether Anthropic can look at your session transcript. Yes ships the whole session, including files a subagent read that you never opened yourself — what the transcript prompt actually uploads breaks down all three answers.

• Add a custom vocabulary for your stack. Library names, API endpoints, and product terms (e.g. "Tanstack", "Supabase", "useMutation") are the words dictation gets wrong. Add them once so they transcribe correctly every time.
• Let Developer Mode handle file names. Don't spell out paths — say the file name naturally and let workspace resolution map it to the real identifier.
• Speak the goal, then the constraints. "Add pagination to the users table — twenty per page, server-side" lands better than narrating implementation detail.
• Dictate in short bursts. Ten-to-thirty-word phrases transcribe most accurately. Pause between thoughts rather than chaining clauses.
• Use a decent microphone. Even a basic USB mic cuts errors on technical terms versus the built-in mic in a noisy room.
• Edit with your hands. Voice is fastest for the first draft of a prompt or comment; fix the last 5% by keyboard. The hold-to-talk hotkey keeps your hands in position.
• Keep proprietary code on-device. If your codebase is confidential, use an on-device tool so spoken file and function names are never transmitted.

Everything above works with any system-wide dictation app. This section is the one place the guide argues for ours, because two of the problems this page describes only have one fix.

Spoken file names that actually resolve. This is the difference between voice being a novelty in Cursor and being the way you work. Say “update the user service and the auth middleware” and Developer Mode matches those against the files in your open workspace, so they land as userService.ts and authMiddleware.ts rather than as four English words your agent then has to guess at. No general speech model does this, because none of them can see your project.

No voice state to get stuck in. Cursor’s voice mode is a mode: you enter it, the transcript hides until you leave it, and the spoken submit keyword that should end it is broken. A held hotkey has none of that shape. The microphone is live while the key is down and dead when it is up, the text appears where you are already typing, and Cursor’s own send key still sends. There is nothing to toggle out of.

And it covers the other four surfaces. Cmd+K inline edits, the chat panel, the code editor itself, the integrated terminal — plus every app outside Cursor, which is where the PR description and the Slack update live.

On the question this page keeps raising: Cursor has not published where its voice transcription runs. Voibe has, because it is a setting you control — on Apple Silicon (M1–M4, macOS 13+) it runs Whisper fully on-device, so the recording never leaves your Mac and only the finished text reaches Cursor. On Windows and Intel Macs it uses a private zero-retention cloud: audio is never stored, sold, or used to train AI. For a tool you point at a proprietary codebase, that is the whole argument.

What it costs. $149 once for a lifetime licence, or $7.50/month, or $59/year (the category comparison is in our dictation app pricing breakdown). Against Wispr Flow’s $144/year subscription — the cloud option in the tools list below — the lifetime licence pays for itself in just over a year and then keeps working. There is a 7-day free trial and no account required to start; full plan details are on Voibe pricing.

Download Voibe for Mac — or grab the Windows installer from getvoibe.com. Then open Composer, hold your key, and say the multi-file change you were about to type out.

Everything above sends your voice into Cursor. The reverse works too — Cursor's agent mode can transcribe audio you already have, which is useful when the spec you need to implement was a call rather than a ticket.

Voibe's speech-to-text API runs a hosted MCP server at https://api.getvoibe.com/mcp; add it as a remote MCP server in Cursor's settings, or have the agent call the three REST endpoints over plain HTTP. Either way you get four tools — create_transcription_job, get_transcript, list_transcripts, get_balance — and a request like “transcribe yesterday's planning call and turn the decisions into a task list” becomes one instruction.

Billing is per second and charged only on a delivered transcript, so an agent's retries cost nothing, and the audio is deleted the moment the text exists. If you would rather not touch a config file, the same server connects in Claude Cowork, Claude desktop or Claude web under Customize › Connectors › Add custom connector. The full walkthrough is in transcribing a Zoom recording.

Two different systems fail in two different ways here, so the first question is which one you are using: Cursor’s built-in voice mode, or a system-wide dictation app typing into Cursor.

Cursor’s voice mode records but no text appears

Usually not a bug. Cursor shows only a waveform while recording and finalizes the transcript when you stop the mic, so nothing appears until you press Cmd+Shift+Space a second time. If you were waiting for a live transcript, that is the whole explanation.

Saying “submit” does nothing

Cursor’s spoken submit keywords are broken as of version 3.11.19 — a regression of an issue that was fixed back in 2.0.32, reported on the Cursor community forum. Stop the recording with Cmd+Shift+Space, then send with Cmd+Return.

Voice mode sends empty or wrong prompts

Users have reported the native voice mode mis-firing on submit keywords and sending empty messages. If you hit this repeatedly, switch to a system-wide dictation tool for prompt entry — you keep voice input without the Agent-prompt-only constraint or the separate voice state.

Text isn't appearing in Cursor at all (system-wide tools)

This is almost always a missing Accessibility permission. Open System Settings > Privacy & Security > Accessibility and confirm your dictation app is enabled. If it is enabled but still not typing, remove it from the list and re-add it to reset the permission — this fixes most cases after an app update.

File names aren't resolving

Confirm Developer Mode is toggled on and that Cursor is the frontmost window with the workspace open. Resolution works against the files in your open project; a file that isn't part of the workspace won't be matched. Adding the term to your custom vocabulary is the fallback.

Technical terms are mis-transcribed

Add the offending library, framework, or API names to your custom vocabulary. General-purpose speech models don't know "Zod" or "Drizzle" until you tell them.

Dictation feels slow

On-device transcription uses the Neural Engine, which is shared with other apps. Close memory-heavy apps, or choose a smaller local model if your Mac has 8 GB of RAM. Transcription is noticeably faster on Apple Silicon than on older hardware.

Cursor is not Mac-only, and neither is dictating into it. Cursor’s download page ships builds for macOS (ARM64, x64 and Universal), Windows (x64 and ARM64, in both system and user installers), and Linux (.deb, RPM and AppImage, ARM64 and x64). The setup above is written for Mac because that is where on-device transcription is available; here is what changes elsewhere.

Windows

Cursor’s native voice mode travels with the app, so the Agent-prompt-only limit and the stop-then-send sequence are the same. For the other surfaces you need a system-wide tool, and you have two realistic options:

• Windows voice typing (Win+H) — free and built in. Put your cursor in any Cursor input, press Win+H, and speak. It needs an internet connection and has no custom vocabulary, so library and project names come out mangled. Our Windows dictation guide covers the setting most people miss.
• Voibe for Windows — a native app (not an Electron port) with the same custom Dictionary as the Mac build. Transcription runs through a private zero-retention cloud rather than on-device, because the fully local mode is Apple Silicon only. See Voibe for Windows and the best AI dictation apps for Windows.

Linux

This is the thin one, honestly. Voibe does not ship a Linux build, and neither do most of the commercial dictation apps in this category — the practical options are the open-source local ones. If you code on Linux, Cursor’s built-in voice mode plus an open-source Whisper front-end is the realistic stack; our open-source dictation roundup covers what actually runs there.

Four practical options for voice in Cursor, with the trade-off that matters for each:

• Voibe — system-wide and on-device, with Developer Mode workspace resolution and custom vocabulary. Works in every Cursor surface and every other app; your audio is never stored, sold, or used to train AI, with a fully on-device mode available (in on-device mode, nothing leaves your Mac). $149 lifetime or $7.50/month, 7-day free trial, no account. Best fit for daily coding-by-voice on Mac. See our best dictation software for developers guide.
• Apple Dictation — free and built into macOS. Fine for occasional use, but it has a session timeout, no custom vocabulary, and no workspace awareness, so technical identifiers need manual fixing.
• Wispr Flow — polished cloud dictation with AI formatting, cross-platform. Capable, but it is cloud-based, so weigh that against dictating proprietary code. $144/year.
• Superwhisper — on-device Whisper modes plus optional cloud LLM cleanup and a flexible per-app mode system. $249.99 lifetime. A strong on-device alternative without dedicated IDE file resolution.
• Cursor native voice — built in and convenient for quick Agent prompts; scoped to the prompt, with the processing-location and reliability caveats above.

Using VS Code rather than the AI-first Cursor? See our companion how to dictate in VS Code guide, which covers Microsoft's on-device VS Code Speech extension. For the broader picture of on-device versus cloud dictation, see our cloud vs local dictation guide and offline dictation privacy on Mac.

Dictation is one piece of a larger agent workflow. For how it fits alongside Claude Code, browser verification with Playwright MCP, and automated pull request review, see our five-tool agentic engineering stack.

Dictating into Cursor's terminal usually means dictating into Claude Code — our guide to dictating in Claude Code covers its built-in /voice mode and the system-wide setup side by side, and the terminal dictation guide handles iTerm2, Warp, and Ghostty (including the Secure Keyboard Entry trap). That side of the workflow also has settings worth two minutes of your time: the Claude Code privacy settings guide covers the training opt-out (on by default for Pro/Max sign-ins) and the single env var that turns off its optional network traffic.

Basics

Can you dictate in Cursor?
Yes. Cursor 2.0 has a native voice mode that fills the Agent prompt, and any system-wide Mac dictation tool types into every Cursor surface — the prompt, Cmd+K, chat, the editor, and the terminal.

Does Cursor have built-in voice input?
Yes, since Cursor 2.0. You hold to speak and the transcript drops into the Agent input. It is scoped mainly to that prompt and does not support @-mentions or model switching by voice, and Cursor has not documented whether transcription is on-device or cloud-based.

Setup

How do I dictate code and comments in the editor?
Use a system-wide tool (Cursor's native voice targets the Agent prompt, not the editor). Put your cursor in the file, hold your hotkey, and speak — comments, docstrings, and Markdown all work, and Developer Mode resolves file names you mention.

What is Developer Mode?
A Voibe setting that detects your open Cursor or VS Code workspace and resolves spoken file names, folder names, and project terms to their exact spelling — so "user service" becomes userService.ts automatically.

Privacy

Is voice dictation safe on a proprietary codebase?
Only if audio is processed on-device. Cloud tools transmit your voice — including file and function names — to a server. On-device tools like Voibe run the model locally so nothing leaves your Mac and dictation works offline.

Workflow

How should I voice-prompt the Agent and Composer?
Speak in short, structured chunks: goal, files, constraints. Use the Five-Part Voice Prompt framework for a repeatable structure, and a system-wide tool so you can dictate the same way into Cmd+K and the editor.

Dictating in Cursor turns the slowest part of AI coding — typing out what you want — into the fastest. Set up a system-wide on-device tool, enable Developer Mode so your file names resolve, and dictate your Cmd+K edits, Composer changes, Agent tasks, comments, and commit messages. Cursor's native voice is a fine shortcut for quick agent prompts, but a system-wide tool covers every surface and keeps your code on your machine.

Voibe is the on-device option built for exactly this: download it free (7-day trial, no account), enable Developer Mode, and dictate your next Composer prompt instead of typing it.

Keep going:

• How to voice-prompt ChatGPT, Claude, and Cursor — the Five-Part Voice Prompt framework
• The voice input workflow — the Talk-Draft-Polish loop for developers and writers
• Best dictation software for developers — the full buyer's view
• Getting started with Voibe — complete setup guide
• How to dictate in VS Code — the editor companion
• How to dictate in Linear & Jira — voice ticket-writing
• How to dictate in Slack — faster messages and replies
• Speech to text on Mac — the category overview
• How to dictate in Microsoft Word — the Dictate button, and the paths that work without it