Source: https://fnflow.app/dictation-codex-cli

# Voice input for Codex CLI and the ChatGPT app
Codex runs in several places on a Mac: the ChatGPT desktop app, a terminal through Codex CLI, and an IDE extension. Voice input works differently in each. The desktop app has its own voice features, the CLI docs don't describe any, and some Mac users have run into a dictation limit error. Here's what we found in OpenAI's documentation and how to dictate into Codex from any of them.

## Where Codex runs, and where voice comes from

OpenAI lists Codex in the ChatGPT desktop app, the mobile app, ChatGPT on the web, Codex CLI, the IDE extension and Codex Cloud. On a Mac you'll most likely use one of the first two: the ChatGPT app, or the CLI in a terminal. The CLI installs with one command and starts in your project folder:

curl -fsSL https://chatgpt.com/codex/install.sh | sh
cd ~/code/my-project
codex
The CLI documentation describes pasting an image into the interactive composer and opening longer prompts in the editor set by VISUAL or EDITOR. It doesn't describe voice input. In the terminal, voice has to come from a tool outside Codex that puts text into the terminal window.

## Voice and dictation in the ChatGPT desktop app

The desktop app has two separate things, and OpenAI's docs keep them apart. ChatGPT Voice is a live conversation. You talk, it answers, and it can start and steer work in Chat, Work and Codex. It's available on Plus, Pro, Business, Edu and Enterprise plans, only one voice chat can be active across the app at a time, and you can set a hotkey in Settings > Voice > Voice chat hotkey. On a Mac you can also turn on Screen context and say "Take a look at this" to share the window in front of you.

Voice dictation is the simpler one. In OpenAI's words, you use it when you only want to turn speech into prompt text before sending it. It can also be switched off. Codex's configuration reference has a key for that:

# ~/.codex/config.toml
[features]
in_app_dictation = false
Setting it to true doesn't bypass account or rollout checks, according to the same page.

## The "Codex dictation usage limit reached" error

A thread on OpenAI's community forum, opened on 16 September 2026, reports this message in the macOS ChatGPT app. The first poster was on the Free plan and saw it after dictating and pressing send, while dictation on the same account worked in the iPad app and on the web. A second user on 29 September saw it after stopping a recording, even though the usage page showed 96% of the monthly limit left. Reinstalling and updating didn't help, and retrying failed recordings from Settings > Voice gave the same error. When we checked, there was no reply from OpenAI in the thread, so we can't tell you the cause. If you hit it, those two reports suggest the web version as a workaround.

## A setup that works in the CLI and the app

A system-wide dictation tool doesn't care which Codex you use. It types into whatever has focus, whether that's the ChatGPT app's composer or a terminal running codex. With FnFlow it looks like this:

- Add project names to Settings → Vocabulary. Whisper takes up to 60 terms as spelling hints, which helps with library and service names it would otherwise spell by ear.
- Click into the Codex composer, in the terminal or the app.
- Hold fn (or right ⌘ if your 🌐 key already does something), describe the task and let go.
- Read the prompt, fix any misheard file or function name and press Enter yourself.

A spoken prompt for Codex might come out like this:

The export job writes empty CSV files when a report has more than ten
thousand rows. I think the pagination loop in the exporter stops after the
first page. Don't change the CSV columns, the finance team imports them.
Add a test with fifteen thousand rows first, then fix it.
For prompts longer than a breath or two, double-tap fn to go hands-free and tap again to finish. Or set EDITOR, open the long-prompt editor from Codex and dictate into that file instead. There's more on structure in [how to dictate a good prompt](dictate-ai-prompts).

## How FnFlow handles Codex

FnFlow transcribes on your Mac with Whisper through whisper.cpp on the GPU, then pastes the text once with ⌘V. On an M4 Pro with large-v3-turbo kept warm, a 7.6-second sentence lands about 1.2 seconds after you let go. Nothing is sent anywhere, so there's no dictation quota and no account to sign in to. After pasting, FnFlow reads the field back and tells you whether the text arrived. If it couldn't paste, the dictation waits in history with its audio.

The limitation is that FnFlow only types. It can't hold a conversation with Codex, answer a question by voice while a task runs or share your screen, which is what ChatGPT Voice is for. If you mostly want to talk to an agent rather than write prompts, that feature is the better fit. The terminal side is covered in more detail in [dictation in Claude Code and the terminal](dictation-claude-code-terminal), and most of it applies to Codex CLI too.

## Questions

**Does Codex CLI have built-in voice input?**
OpenAI's Codex CLI page, which we read on 18 October 2026, doesn't describe one. It covers pasting images into the composer and opening longer prompts in your VISUAL or EDITOR editor. For voice in the terminal you need a dictation tool that types into the terminal window.

**What's the difference between ChatGPT Voice and dictation?**
OpenAI's docs describe Voice as a live conversation that can direct work in the desktop app, available on Plus, Pro, Business, Edu and Enterprise plans. Dictation only turns your speech into prompt text before you send it.

**Why do I see "Codex dictation usage limit reached"?**
We don't know. A thread on OpenAI's community forum describes the error in the macOS ChatGPT app, including from a user whose usage page showed 96% of the monthly limit left. Dictation worked for the same accounts on the web and on iPhone or iPad. OpenAI hadn't replied in the thread when we checked.

**Does FnFlow have a usage limit?**
No. FnFlow is free with no word limit, no account and no API keys. Recognition runs on your Mac, so there's no server-side quota to run out of.
