← All posts

Dictate prompts into Codex CLI on your Mac

To dictate prompts into Codex CLI on a Mac, run codex in your project, click into its prompt, hold the hotkey of a dictation app that types at the cursor, talk through the task, and release. The text lands in the Codex composer and you press Enter.

That part takes a minute to set up. What takes longer to learn is the rest: which hotkey not to use, how to speak a task Codex can run with, and what to do when the words do not show up. This post covers all three.

Does Codex CLI have its own voice input?

Partly, and it keeps changing. Codex CLI has had experimental voice features during 2026, some of them were removed again, and newer builds have a Voice mode that talks back. What it does not have is plain dictation you can rely on: a request for “native voice dictation in Codex CLI” was opened on the openai/codex tracker on 4 October 2026 and closed as a duplicate of an older request.

So most people who talk to Codex use a system wide dictation app. It types at the cursor in any terminal, it works the same in Codex, Claude Code and your editor, and it does not depend on which Codex version you installed this week.

Codex details checked on 8 October 2026. Sources are linked at the end.

What to set up before your first dictated prompt

You need three things.

  1. A dictation app that types at the cursor. macOS Dictation does this, and so does Akova, a Mac dictation app for coding agents. Akova runs on your Mac with on-device Whisper, needs macOS 14 Sonoma or later on Apple Silicon, and has a Prompt mode that turns a spoken ramble into a structured prompt before it types.
  2. Your project words in custom vocabulary. Codex prompts are full of file names and library names that a speech model has never heard. Add them once.
  3. A hotkey that Codex does not want. This is the one people skip, and the next section is why it matters.

Here is a vocabulary list for a typical web repo:

AGENTS.md
config.toml
pnpm
Vitest
useCartStore
payments-service

Without it, “Vitest” comes out as “vi test” and “pnpm” as “PNPM” or worse. Codex can usually guess, but every guess costs you a reread.

Pick a hotkey that does not start Codex Voice mode

On 29 September 2026 a user reported on the OpenWhispr tracker that pressing their dictation shortcut while Codex CLI 0.158 was focused in Terminal started Codex’s own Voice mode instead of their dictation app. It happened on the Fn+F5 dictation key and again after they moved the shortcut to Fn+F11.

You do not need to know exactly why to avoid it. Choose a hotkey that:

  • is not the Fn or globe key alone and not the F5 microphone key,
  • is not Space, because terminal apps read Space as text and some voice features have used hold Space to talk,
  • you can hold with one hand without looking, because you will press it all day.

A right side modifier, such as right Option or right Command, is a good default. In Akova you set it in Settings. If your dictation hotkey ever opens a Codex screen instead of recording, change the hotkey first and debug second.

Dictate a Codex task, step by step

This is the loop we use.

  1. cd into the repo and run codex.
  2. Click into the prompt so the cursor is in the composer.
  3. Hold your hotkey and say the task the way you would explain it to a colleague. Do not compose. Half sentences and “no wait, I mean” are fine.
  4. Release. With Prompt mode on, Akova cleans the ramble into a structured prompt and types it at the cursor.
  5. Read it, fix a word if you need to, and press Enter.

Here is the kind of thing you might say into Codex for a small bug:

okay so the cart total is wrong when you apply a coupon and then remove an
item, it keeps the old discount, I think it's in useCartStore, the
recalculate function, um, don't touch the checkout API, and add a Vitest case
for remove after coupon, then run the tests

You said the goal, the place, the constraint and the check. Prompt mode’s job is to put them in order and drop the filler, so Codex reads a clean task.

How to speak a prompt Codex can act on

Talking makes prompts longer, which mostly helps: Codex works better with context, and you give more of it out loud than you would ever type. Longer is not clearer, though, so four habits pay off.

Lead with the outcome. “The cart total should drop the discount when the coupon item is removed” is a better first sentence than a tour of the code.

Name files and functions. “In useCartStore, the recalculate function” gives Codex a starting point. Custom vocabulary makes sure it is spelled the way it is in the repo.

Say what must not change. Spoken constraints are the ones people forget to type. If your repo has an AGENTS.md, Codex already reads your standing rules from it, so you only need to say what is special about this task.

Say how to verify. “Run the tests” or “add a test for the empty cart”. Codex will do it if you ask, and it is easy to ask out loud.

The same habits work in Claude Code. If you use both agents, our guide to dictating prompts into Claude Code has the matching setup for that terminal.

When dictated text does not appear in Codex

Two problems come up in the Codex issue tracker and in forums.

The words appear only after you press a key. A macOS user reported on the openai/codex tracker that dictated text, mixed Chinese and English, did not show in the Codex input line until they pressed Enter or another key, starting from the second turn of a session. They saw it in Cursor’s terminal and in Antigravity. If you hit this, do not press Enter blind, because that sends a prompt you have not read. The reporter saw the text appear after other keys too, so try one that does not send, such as an arrow key, then read and send. Updating Codex is worth a try, since the report was tied to specific updates.

The hotkey opens something else. That is the Voice mode collision above. Change the dictation hotkey.

If neither applies and nothing types at all, check that your dictation app is allowed under Accessibility in System Settings, Privacy and Security. Apps that type into other apps usually need that permission. If it is, and dictation works everywhere except your terminal, read why dictation stops typing in the Mac terminal: the usual cause is Secure Keyboard Entry.

FAQ

Can I use the built in macOS Dictation with Codex CLI?

Yes. macOS Dictation types into Terminal, so it works in Codex. It types exactly what you said, without your project vocabulary and without turning a ramble into a prompt.

Does my audio go to the cloud?

Not with Akova by default. It transcribes with on-device Whisper models and Prompt mode runs locally too, so it works offline. For the trade offs, read on-device vs cloud dictation for Mac developers.

Is Prompt mode free?

It is a Lifetime feature. Every Pro feature, Prompt mode included, is free for 7 days after your first launch, with no card and no account. After that, Clean up, custom vocabulary and audio file transcription stay in the free tier. Lifetime adds Prompt mode, more modes, review before pasting and Voice for Agents (beta). See pricing for the full list.

Try it on your next Codex task

Add five words from your current repo to custom vocabulary, set a right side modifier as the hotkey, and dictate the next task you were about to type into Codex. Download Akova free, no account and no card.

Sources

Checked on 8 October 2026.