Voice dictation for Codex: say the whole task for Codex CLI, the IDE extension or the app

Say the whole task to Codex.

What Codex offers for voice today

We checked OpenAI's docs, the Codex CLI release notes and the openai/codex GitHub repository on October 18, 2026. Here is what they say.

• In the ChatGPT desktop app, which runs Codex, OpenAI's docs say you can select Dictate in the composer and start talking. Your speech is transcribed into the composer so you can review and edit it before sending. If you work in the app, start there.
• In Codex CLI, /voice starts a spoken conversation with Codex. The CLI release notes added experimental /voice conversations in version 0.155.0 and turned voice conversations on by default in 0.156.0, with F8 to toggle them and a /voice settings picker.
• Codex CLI does not offer one-way dictation into the prompt. An experimental hold-Space transcription added in 0.105.0 was removed in 0.118.0. On April 2, 2026, an OpenAI maintainer wrote on GitHub that voice transcription is fully supported in the desktop app but not available in the CLI. On October 12, 2026, a maintainer wrote that OpenAI has no plans to implement one-way voice transcription in the CLI.

This can change with any release. Check OpenAI's docs and the Codex CLI release notes for the current state.

Codex CLI users have asked for prompt dictation

Two openai/codex GitHub issues ask for dictation in Codex CLI. Issue #16404 (April 2026, 20 thumbs-up when we checked) reports that the TUI voice transcription flow was removed in 0.118.0 and asks for a supported CLI dictation path; it was closed as a duplicate. Issue #46698 (September 2026) asks to bring composer dictation to the CLI so speech lands as an editable draft the user reviews and submits; it was closed as not planned. We checked both on October 18, 2026.

Dictate, review, then copy and paste into Codex

1. Dictate the task. Hold Right Alt on Windows, Right Option on Mac or Left Ctrl + Left Alt + Space on Linux, and talk into a notes window or chat box.
2. Review it. Read the text, fix names and paths, or select it and run Transforms > Create an AI prompt.
3. Copy and paste it into Codex. In Codex CLI, paste into the prompt in your terminal. In the Codex IDE extension or the app, paste into the input box.
4. Send it yourself. Nothing reaches Codex until you press Enter.

When a separate dictation app makes sense

• You work in Codex CLI and want to write the prompt, not talk it through. /voice is a conversation with the agent. Whispek gives you plain text you can read and edit before Codex sees it.
• You want to read the prompt before it is sent. With Whispek nothing is sent until you press Enter in Codex.
• You work over SSH or in a remote shell. Whispek runs on your own computer, so you dictate locally, then copy and paste the reviewed text into the remote session.
• Codex is one of several places you write. The same Whispek shortcut works in Cursor, browser AI chats, GitHub, Slack and email.

Codex's built-in voice vs Whispek

How to dictate a Codex CLI prompt with Whispek

1. Open somewhere to draft. A notes window, a chat box or your editor's scratch file. Say the context, goal, constraints and done-when check. This works the same for the Codex IDE extension: draft first, then copy and paste into its input box.
2. Press your Voice input shortcut and describe the task. Right Alt on Windows, Right Option on macOS, Left Ctrl + Left Alt + Space on Linux; you can change the Voice input shortcut on the Home screen. Turn on Settings > Voice & translation > Live transcription display to see the transcript above the capsule while you talk.
3. Structure a rough prompt if needed. Select the dictated text, open the Transforms window and choose Create an AI prompt, then review the result. For your words as spoken, choose Original in Style > Cleanup before dictating.
4. Copy and paste it into Codex, add exact context, then send. Add files with @ in Codex, paste exact errors and paths, read the prompt once and press Enter. If the text did not appear where you expected, copy it from History.

A spoken structure for a Codex task

• Goal: "The CSV export times out for accounts with more than 50,000 rows."
• Context: "The export is built inside the request handler in the exports module."
• Constraints: "Keep the column order and the file name format. Don't add a new dependency."
• Done when: "An export of 100,000 rows finishes and the existing export tests pass."
• Plan first: "Show me your plan before you change any files."

Paste the error and the exact file path yourself rather than reading them out. Do not dictate shell commands to run directly.

Using Claude Code too?

Claude Code has its own /voice dictation, with limits on sign-in methods and remote sessions. See voice dictation for Claude Code for when its built-in option fits and when to draft with Whispek instead.

Keep names right and reuse what works

• Put your stack in Dictionary once, for example Kubernetes, PostgreSQL, Tauri, useEffect and pnpm, and they're in your dictionary on every computer you sign in to, so Whispek writes them your way. Import a whole list with Dictionary > More actions > Import terms.
• Save standing instructions as Snippets, such as "Show me the plan first. Run the tests before you finish."
• Find earlier prompts in History and copy them again.
• Create a New transform with your own Instructions, for example to turn a rambling request into goal, context, constraints and done-when.

What Whispek does not do in Codex

• It is not an OpenAI product or a Codex plugin.
• It does not read your repository, run commands, approve changes or send prompts.
• It does not talk back. For a spoken conversation with Codex, use /voice in Codex CLI or ChatGPT Voice in the desktop app.

Pricing

Whispek is free every 7 days. Pro is unlimited for $8.25/month, billed yearly ($99), or $12 month to month. Prices exclude tax.

Your next Codex task, out loud

Install Whispek, open a notes window next to your terminal and say the task you would normally cut down to one line.

Official sources