Codex voice mode: talk to Codex and hear its answers
SpeakApp reads each Codex reply aloud, one after another, and you say “ok” to answer the session that spoke. It works with the Codex app for Windows, for listening and for answering by voice. Dictation types your words at the cursor and is a separate function from the reading loop. Keep keyboard review for code, diffs and output that needs visual inspection.
What does a Codex voice workflow look like in practice?
Your agents' replies are read aloud one after another, as soon as they are ready. You say “ok”, speak your answer, then say “enter”, and your message goes straight into the session that spoke, without hunting for its window, touching the keyboard or looking at the screen.
Suppose a Codex task is investigating a failing test, another Codex task is reviewing a proposed refactor, and a ChatGPT chat in Chrome is helping you draft an explanation for a colleague. These are example tasks, not a requirement for how you organize your work.
The test investigation finishes while you are thinking about the refactor. Its reply joins the queue and is read aloud. You say “ok”, ask the agent to check the relevant test setup, then say “enter”. Your message goes to that conversation rather than whichever window you last touched.
That is the useful distinction: entering words by voice is one task, and keeping those words attached to the right conversation is another.
Why does the reply loop matter more than dictation?
Dictation solves the input step. A multi-session workflow also needs a way to hear completed answers, choose what to handle, and reply in the right place.
In the example, a perfectly transcribed instruction still creates a problem if it lands in the refactor conversation instead of the test investigation. The wording is correct. The destination is wrong. Treat destination control as a separate requirement from speech recognition.
The listening queue also gives pending conversations a place to wait. A toast, the panel SpeakApp shows above your windows, offers read, skip, later, reply and open, while the widget's circle counts what is still waiting. You do not have to treat every finished answer as the next task you must handle.
SpeakApp is for people who run several AI sessions at once, and developers running several coding agents are a common case. It reads each reply aloud in turn and sends your spoken answer back to the session that spoke.
The trade-off is attention. Hearing an answer does not make its reasoning correct, and a spoken summary does not replace inspecting a patch. Keep the listening loop for conversational progress, then return to the screen when the work demands exact visual review.
How do the Codex voice options compare at a glance?
Choose between a connected listening-and-reply loop, voice input alone, and manual interaction. These are different workflows, not interchangeable versions of the same feature.
| Option | Best for | Main function | Where you still do the work |
|---|---|---|---|
| SpeakApp with the Codex app for Windows | Several agent sessions with spoken follow-ups | Queues completed replies and routes your spoken response to the right session | Reviewing code and other visual details |
| Dictation into a prompt field | Speaking a prompt instead of typing it | Converts speech into text | Selecting the conversation and handling its answer |
| Keyboard and screen only | Deliberate, screen-focused review | You read the reply on screen, find the right window and type the next instruction | Monitoring and switching between conversations |
Compare the whole cycle rather than asking only whether a microphone can enter text. Can you hear the response? Can you defer it? Can you reply to that conversation without selecting another window?
This page covers what SpeakApp adds around the Codex app. It does not describe any feature built into Codex itself.
When should you choose the queued voice loop?
Choose the queued voice loop when several supported sessions are active and you want to handle their replies without repeatedly finding each conversation. The distinguishing feature is the connection between the answer you hear and the session that receives your response.
In the running example, the refactor agent can finish while you are answering the test investigation. Its conversation waits in the queue. You can handle it afterward instead of abandoning the current exchange to find another window.
Where this option shines
- Completed answers are read aloud one after another.
- The listening queue keeps waiting conversations visible.
- Your spoken answer returns to the conversation selected for reply.
- Dictation into other text fields is available as a separate function.
Where this option falls short
- It follows the apps listed in the section on supported sessions below, so check that list before you rely on it.
- Sessions running in the cloud are not read.
- Audio does not replace checking code, file names or a detailed diff on screen.
- Support for listening does not automatically mean support for replying by voice in every application.
| Dimension | Queued voice loop | Keyboard and screen only |
|---|---|---|
| Receiving an answer | Hear the queued reply | Read the selected conversation |
| Choosing what waits | Use queue and toast actions | Manage attention by switching conversations |
| Sending a follow-up | Speak back to the right session | Find the window and type |
| Inspecting a patch | Return to visual review | Stay in visual review |
Best for: people who run several supported agent sessions, such as developers running several coding agents.
Verdict: Choose the queued loop when listening and responding across sessions is the task. Keep visual review in the workflow.
When is dictation alone enough for Codex?
Dictation is enough when your main problem is entering a prompt, not hearing answers or managing several conversations. Select the intended field, speak your instruction, and review the text before sending it.
In the example, you could dictate a test-investigation request and then read the result yourself. That is a reasonable choice if you want to remain focused on the editor and do not need responses read aloud.
Where dictation shines
- You can speak the explanation you would otherwise type.
- You keep control of the destination by selecting the prompt field yourself.
- It fits a workflow where you already read every answer on screen.
Where dictation falls short
- Text entry alone does not provide a queue of completed agent replies.
- You still need a separate way to hear an answer.
- You still select the conversation that receives your follow-up.
| Dimension | Dictation alone | Keyboard and screen only |
|---|---|---|
| Prompt entry | Speak into the selected field | Type into the selected field |
| Answer review | Handle separately | Read on screen |
| Session selection | Choose the destination yourself | Choose the destination yourself |
Best for: a screen-focused workflow where speaking replaces typing but does not replace session management.
Verdict: Choose dictation for input. Skip it as the complete solution when you also need queued answers and session-aware replies.
When should you stay with the keyboard?
Stay with the keyboard when the immediate task depends on seeing exact text and controlling each edit. Reading a diff, checking punctuation in a command, or comparing file paths are reasons to keep the screen central.
The manual workflow is also a useful baseline for your decision. Try the same task with and without the voice loop, then judge whether the interaction fits how you work. Do not substitute an assumed time saving for that check.
Where manual interaction shines
- You can inspect the answer and compose your response in the same visual context.
- You can select precise text for a correction.
- You control which conversation has your attention.
Where manual interaction falls short
- You must find and open the conversation you want to handle.
- You must read the answer before responding to its contents.
- You remain responsible for tracking the other waiting sessions.
Best for: detailed code review and tasks where exact visual inspection is the main work.
Verdict: Hold onto the keyboard for precision work. Add voice around it rather than treating voice as a replacement for every interaction.
How do you run the listening-and-reply cycle?
Start with supported sessions and a clear task in each one. Then handle a completed conversation through the listening queue instead of choosing a destination from memory.
Use the same sequence for the test investigation, refactor review and colleague explanation:
- Queue replies. Let completed answers wait for listening.
- Hear an answer. Listen to the conversation being presented.
- Say ok. Start a spoken reply to the conversation shown in the toast.
- Reply by voice, then say enter. Give the follow-up for that conversation, and “enter” sends it into the session that spoke.
If an answer calls for a patch review, open the conversation and inspect it. If it does not deserve your attention yet, use later. If it is not useful to hear, skip it. The point is to choose what you handle, not to listen to everything simply because it is ready.
Which Codex, Claude and ChatGPT sessions are supported?
Listening and answering are two separate lists. SpeakApp answers by voice in the Codex app for Windows and in Claude Desktop (the Code tab and local Cowork), and on ChatGPT on the web in Chrome through the SpeakApp extension.
Listening also works in Cursor and in VS Code, and on Claude.ai on the web through the SpeakApp extension. Do not infer reply support from that listening list. Sessions that run in the cloud are not read.
That boundary matters in the example. The Codex tasks have to run in the Codex app, and the ChatGPT chat has to be open in Chrome with the extension.
Dictation, which types at your cursor, is a separate function from this reading loop. So is Hey Speak, which answers a question out loud or passes it to one of the AIs you already use.
Why move beyond manual switching or a single voice conversation?
Move beyond manual switching when selecting the next session has become a separate job from deciding what to tell the agent. Keep the manual workflow when reading and typing are already the right fit.
A built-in voice mode, such as the ones in ChatGPT, Claude or Gemini, keeps you in one conversation. Before you treat one as a substitute, ask whether it handles the exact application, local session and cross-session reply workflow you use.
Use a concrete acceptance test: let the test investigation finish, leave the refactor running, and check whether you can hear the completed answer and respond to that specific task. Then inspect the destination. Correct words in the wrong conversation fail the test.
What should you check before choosing?
Check the destination of the reply before judging the convenience of speaking. That is the part of a multi-session workflow that a microphone test alone misses.
Your practical checklist is short:
- Confirm the application is covered for the action you need.
- Separate local sessions from cloud-running sessions.
- Try hearing an answer while another task remains active.
- Confirm the follow-up reaches the intended conversation.
- Keep a visual check for changes that require exact review.
The useful result is not a longer spoken exchange. It is a complete loop that leaves you knowing which agent answered and which agent received your next instruction.
Before your first session, give each agent a distinct task. It makes the destination check concrete: a reply about the failing test belongs with the test investigation, not the refactor.
Related guides
Frequently asked questions
How can I talk to Codex and hear its answers?
Use SpeakApp with the Codex app for Windows. Each finished reply is read aloud, one after another. Say “ok”, speak your answer, then say “enter”: your message goes into the session that spoke.
Is dictation the same as hearing Codex's answers?
No. Dictation types your words where your cursor is. The reading loop reads each finished reply aloud and sends your spoken follow-up to the session that spoke.
Can I manage Codex and ChatGPT sessions together?
Yes. Replies from the Codex app for Windows and from ChatGPT on the web in Chrome (through the SpeakApp extension) wait in the same queue, and you answer each one in its own session.
Can I listen to sessions running in the cloud?
No. A conversation that runs remotely in the cloud leaves no trace on your disk, so SpeakApp can neither read it nor reply to it.
Should I stop reviewing code on screen?
No. Keep visual review for code, diffs and other details that need exact inspection. Voice changes how you receive answers and give follow-ups, not whether you verify the work.
Try SpeakApp
SpeakApp is available on Windows, coming soon to Mac. Download SpeakApp or join the Mac waiting list. To see the whole product first: What is SpeakApp?