Voice mode in Codex: how to enable /voice and what it can do

Stable Codex CLI 0.155.0 was released on September 17, 2026, followed by the 0.155.1 fix the next day. Version 0.155.0 introduced experimental voice conversations with the agent in the terminal interface: you can describe a task aloud and hear the response. Use the /voice command to start.

The feature is disabled by default and is not supported everywhere. To enable it, open /experimental, select Voice conversations and restart Codex. The source description was checked against the release notes, change #43581 and the Codex source code on September 19, 2026. The source material did not run an actual voice conversation: the mechanics are described from primary sources, with no firsthand test here.

Additional instructions are available in the Ailoot documentation.

What voice features were added in Codex 0.155?#

There is a closed discussion in the repository about the removal of /realtime. Version 0.155.0 brings voice back to the terminal in a new form: a single /voice command with a live transcript, recording indicators and microphone controls.

Conditions and limitations:

• the feature is experimental and disabled by default;
• it works on supported builds — OpenAI does not name an exact list of platforms in the notes, and support is checked at startup;
• the description of change #43581 mentions local WebRTC audio and signaling through app-server. That alone does not mean the conversation stays entirely on your machine; the notes do not detail the exact processing route.

Version 0.155.0 also added live reasoning summaries in the status line, task management in the agent overview (hide, archive and delete), approval of requests to MCP servers (connected external tools) through Touch ID on Mac, a configurable update schedule for the background service, and retrieval of Amazon Bedrock credentials from external commands. Version 0.155.1 is a targeted fix: reasoning summaries are once again disabled by default for new local sessions because some providers were rejecting those requests.

How do you update Codex to 0.155?#

First, check the installed version:

codex --version

If the version is below 0.155.0, update it using the same method you used to install it:

# If you installed through npm
npm update -g @openai/codex

# If you installed through Homebrew
brew upgrade --cask codex

If you installed with the official script from the website, run it again:

# macOS or Linux
curl -fsSL https://chatgpt.com/codex/install.sh | sh

On Windows, use the same PowerShell installer as for the first installation. If you have not installed Codex yet or are unsure how to sign in, start with the guide Codex from OpenAI: what it is, how to install it and sign in.

How do you enable voice mode?#

Steps:

1. In a running Codex session, enter /experimental.
2. Find Voice conversations in the list of experimental features (its description is “Talk with Codex using /voice”) and enable it.
3. Exit Codex completely and launch it again. The feature announcement says: “Restart Codex after enabling, then use /voice”.
4. Enter /voice — a voice conversation will begin.

You will need a microphone and operating-system permission to use it. On Mac, the system will ask for access when recording starts for the first time; on Windows and Linux, check that the microphone is selected as the default device. If Voice conversations is missing from /experimental, your build most likely does not support the feature — update to the latest version and check again.

What can a voice conversation do?#

According to the description of change #43581 in the repository, the conversation works as follows:

• Real-time transcript. Your words and the agent's replies appear on screen as they are spoken, alongside the conversation status and microphone and speaker audio levels. The text is not lost when switching between conversations.
• Not everything is spoken aloud. The agent's final replies in a voice conversation are spoken. Reasoning and intermediate commentary are hidden. Replies received as text are not repeated aloud.
• Mouse-free controls. /voice mute mutes the microphone (the keyboard shortcut for this is configurable), and /voice stop ends the conversation.
• Reduced logging. The text spoken aloud and voice-session event data were removed from receipt and debug logs.

Audio is processed by OpenAI's service. Do not share sensitive information in a voice conversation — client data or figures that must not leave your machine.

When is voice useful, and when is text better?#

A few scenarios for /voice:

• Dictate the whole task. If speaking is faster than typing for you, this is a way to provide lengthy context: the background, what you have already tried and where not to go. It is the same approach as dictating tasks in other tools, now within Codex itself.
• Ask as you go. While the agent is working, you can ask aloud “what are you doing now?” or “why did you choose that option?” without switching to the keyboard.
• Discuss the plan before coding. Talking through three possible solutions and hearing an analysis of them makes for a live discussion.

It is more reliable to paste exact commands, file paths, code snippets and error messages as text. If you only need quick dictation without a conversation, a local option such as Whisper does this for free without sending audio outside your machine; that option is covered in the article on free voice input with Whisper.

What are the limitations, and what should you check?#

Check these limitations before using it:

• Experimental status. The release notes state this explicitly. Commands and behavior may change; if something does not work as expected, consult the notes for the latest version.
• Not all builds. OpenAI uses the wording “supported builds” but does not publish a list of platforms in the release notes — support is determined at startup. If the option is missing from /experimental, check your version and whether your build supports the feature.
• Plans are not disclosed. The release notes do not say which ChatGPT subscriptions provide access to voice conversations. If /voice does not start even though the feature is enabled, check availability in your plan's documentation.
• Audio leaves your machine. The conversation does not stay entirely on your machine; OpenAI does not detail the exact processing route in the notes. The implementation description mentions local WebRTC audio and signaling through app-server. Before discussing work data, check whether transferring it is permitted.
• Enabled by default in 0.156. The code already contains a change that moves voice conversations from experimental to stable and enables them by default; it was included in prerelease 0.156.0-alpha builds, but not in stable 0.155. Judging from the code, once 0.156 is released, separate activation through /experimental will no longer be necessary.

If you are choosing between terminal agents, voice is only one criterion. For an overall comparison of capabilities, see the Codex or Claude Code overview.

Brief recap#

Codex 0.155.0 brought voice conversations with the agent to the terminal: enable the mode through /experimental, use /voice after restarting, and control it with /voice mute and /voice stop. The agent speaks only final replies in a voice conversation, reasoning is hidden, and a live transcript appears on screen. The feature is experimental and is not supported on every build, and audio leaves your machine; prerelease builds in the 0.156 series are already moving the mode to stable status by default. Update with the command for your installation method, enable it and try it on your own task.

When you are ready, explore Ailoot.