Codex CLI: Audio Inputs and Realtime V3 Voice Conversations
CodexCodex CLI 0.145.0 added support for audio inputs and tool outputs, including common local audio formats, and introduced a new streaming realtime V3 conversation mode. This builds on Codex's existing voice capabilities, letting developers speak instructions to Codex and receive audio back from tools, rather than being limited to text-only exchanges. The update also lets developers freely mix voice and typed input within the same session.
Key Takeaways
- Codex CLI now supports audio inputs and tool outputs, extending voice beyond simple dictation.
- The new streaming realtime V3 mode is the third major iteration of Codex's voice conversation architecture in 2026.
- Developers can mix voice and typed input within the same session for a hybrid workflow.
- Tools invoked during a session can now return audio as part of their output, not just receive spoken input.
- This builds on Codex CLI's WebRTC-based realtime voice stack introduced earlier in the year.
- The feature targets developers who prefer speaking high-level instructions while typing precise details like file paths.
Sources & Mentions
3 external resources covering this update
Voice Gets a Third Iteration
Codex CLI has offered a realtime voice mode since earlier in 2026, evolving from push-to-talk dictation to a WebRTC-based two-way conversation system. With version 0.145.0, OpenAI introduced streaming realtime V3 conversations, along with support for audio inputs and tool outputs in common local audio formats.
Talking to Tools, Not Just Codex
Previously, voice input in Codex was primarily used to dictate instructions that were transcribed into text before entering the agentic loop. The new audio tool outputs extend this further, allowing tools invoked during a session to return audio content as part of their output, not just receive it as input.
Mixed Voice and Text Workflows
The update supports freely mixing voice and typed input within a single session. In practice, this lets developers speak high-level instructions and then switch to typing specific file paths or code snippets that are awkward to dictate, combining the speed of voice with the precision of text where each is best suited.