OpenAI's Codex Voice: Real-Time Voice That Runs Your Computer

TL;DR

OpenAI put real-time voice directly inside Codex — not dictation, but a live voice agent that can see your screen, use the skills you've already built, control an in-app browser, and spin up parallel background tasks while you keep talking. In one 16-minute first test, Riley Brown had it triage a support inbox and draft 11 replies, summarize the issues into a public Google Doc and text it to someone, read/edit an Excalidraw board, run background Notion research, and build three iOS apps at once by voice — then did it all again remotely from the ChatGPT phone app controlling his Mac. The interesting part for builders isn't the "Jarvis" wow — it's the orchestration model (one live conversation coordinating many background agents) and the hard safeguard boundary baked in.

Key takeaways

• Voice ≠ dictation. It's a live agent ("start new voice chat") that acts while you talk — you can even mute and keep working as it runs.
• It uses YOUR skills. Riley's point: "I've set up so many skills… now I can use them in real-time voice." Six months of skill-building suddenly has a voice front-end.
• Parallel background tasks are the real unlock. Mid-conversation it spins up separate chats (email summary, Notion research, 3 iOS builds) and coordinates them while the live chat stays open.
• Remote from your phone. ChatGPT iOS app → paired remote access to your Mac; create and check tasks from anywhere (Android coming).
• Safeguards can't be skilled around. Live voice won't send messages, spend money, change permissions, swap its model, or touch private data unless you explicitly ask — and "a skill can't give itself permission" to bypass that.
• Still early. ~30–45 min of first-time testing; a real best-practices workflow video is promised later. Treat this as "what's now possible," not a proven daily driver.

---

What it actually did (one 16-min session)

1. Support triage → drafts. "Find complaints, draft responses, open them in the Codex browser." → opened 12 threads, saved 11 drafts.
2. Summarize → share → notify. Turned the issues into a public Google Doc (names stripped) and texted the link to a contact, then appended a follow-up line — all by voice.

3. See & edit a canvas. Read his Excalidraw board, then suggested three more capabilities and added nodes to it.

4. Background research. Kicked off Notion research in a separate task (his tab/hyperlink formatting) while the live chat continued.
5. Build 3 iOS apps at once. An Apple-Notes clone in the in-app simulator, plus a Notion clone and a Trello clone — branched into parallel build tasks — then styled by voice ("light mode," "higher-quality liquid-glass tab bar").
6. Visual inspiration on tap. Pulled "cool note UIs" into browser tabs from Google Images while it worked.

---

The orchestration model (the part worth stealing)

The mental model isn't "a voice assistant" — it's one live coordinator conversation fanning out to many background agents:

flowchart TD
    V["🎙️ You (voice)<br/>Mac mic OR ChatGPT phone app"] --> O["Codex real-time voice<br/>(live coordinator — stays open)"]
    O --> S["Your skills<br/>(Gmail, build-iOS-app, …)"]
    O --> B["In-app browser<br/>(read/act on screen)"]
    O -.->|"spins up in parallel"| T1["Task: summarize support inbox"]
    O -.-> T2["Task: Notion research"]
    O -.-> T3["Task: build 3 iOS apps"]
    T1 & T2 & T3 -->|"report back"| O

This is the "software factory" pattern with a voice conductor: you stay in one conversation and dispatch/verify parallel work — exactly the shape worth designing agent systems around.

The safeguard boundary (and why it matters)

Notably, the live voice mode is intentionally constrained, and crucially skills cannot override it:

flowchart LR
    subgraph CAN["✅ Does freely"]
      A["Read your screen / browser"]
      B["Draft texts, emails, Slack, posts"]
      C["Build docs / code in background"]
      D["Spin up background tasks"]
    end
    subgraph NEEDS["🔒 Only if you explicitly ask"]
      E["Send messages"]
      F["Spend money"]
      G["Change permissions"]
      H["Access private data"]
      I["Swap its own model"]
    end

"A skill can't give itself permission to send things or make consequential changes on its own." For anyone wiring up autonomous agents (me included), that separation — draft/build freely, but gate every consequential/outbound action behind an explicit human ask — is the right default. It lines up exactly with keeping sends, spends, and secrets human-gated.

---

This was a raw first-look, not a workflow guide — Riley plans a deeper "how I actually use this in my business" video next.