Codex Desktop: add a read-aloud / replay audio button for assistant responses
Feature request
Add a read aloud / replay audio button to assistant responses in the Codex desktop app, similar to the speaker/replay behavior available on chatgpt.com.
Why
Codex Desktop is increasingly used for long-running, conversational coding work. Being able to replay an assistant response as voice would make the desktop app easier to use while multitasking, reviewing long explanations, or working in a hands-busy flow.
This would also improve accessibility and parity with ChatGPT's web UI.
Proposed UX
For each assistant response in Codex Desktop:
- Show a small speaker/play button in the response action row.
- Clicking it reads the response aloud.
- While playing, show pause/stop/replay state.
- The action should work on already-completed responses, not only during streaming.
- Ideally, preserve the same behavior/style as chatgpt.com so users do not have to learn a new control.
Nice-to-have behavior
- Respect the user's ChatGPT/Codex voice settings if available.
- Provide a simple fallback OS TTS voice if cloud voice playback is unavailable.
- Avoid reading hidden metadata, tool JSON, or internal reasoning; read only the visible assistant answer unless the user explicitly expands/selects another item.
- If a response contains code blocks, either skip them by default or use a readable summary mode.
Context
The desktop app already has rich response actions and long-form coding conversations. Voice replay would be a small but high-leverage quality-of-life feature, especially for users who keep Codex open as an active desktop companion while coding.
5 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
This would also be great in the mobile app, for using Codex on the go!
would be great!
There is an important accessibility distinction between Read aloud, Rephrase, and Repeat that would make this substantially better for Voice users:
turn/start/turn/steer, rewrite the wording, or appear as “Thinking.”This is especially useful for ADHD, auditory-processing differences, distraction, and hands-busy workflows: it should behave like replaying a podcast segment, not asking an intelligence to reconstruct what it said.
A tap/keyboard/accessibility action can guarantee no new model generation or Codex tokens. A spoken “say again” can only carry a literal zero-usage guarantee if the command is recognized locally before upstream processing; otherwise the honest promise is “replays without rerunning Codex,” because speech recognition may still use part of the separate Voice allowance.
Suggested V1 acceptance criteria: cached replay emits no model/turn request, starts within 100 ms on supported hardware, is barge-in interruptible, does not alter conversation history, and never labels partial audio as a complete replay.
This should cover iPhone and iPad too, not only desktop.
My use case is reviewing Codex while walking, on e scooter/motorbike. I want to listen to a completed answer, then reply without stopping to read the screen.
The useful version:
This should be TTS playback of the completed answer, not a new model turn or Voice mode. Highlighting should not block the MVP.