Direct Answer
ChatGPT Voice can now act as the live control layer for Work and Codex in the desktop app. You can talk through a goal, interrupt or redirect the conversation, ask it to start background tasks, check their progress, and coordinate agents that use the tools and permissions already available in Work or Codex.
Calling the feature "Codex Voice" or "Jarvis" is understandable, but technically incomplete. It is ChatGPT Voice powered by GPT-Live, connected to the selected Work or Codex environment. The voice model handles the natural conversation. Work or Codex handles the actual task context, browser, files, code, tools, and delegated work.
Credits: The hands-on workflow test is by Riley Brown. Follow @rileybrown on X. The second embedded video and product documentation are from OpenAI.
Source Note
Riley's video is a first-session demonstration recorded after the feature launched. It is strong evidence of what worked in his configured environment, but not proof that every account, connector, operating system, or workflow will behave identically.
I checked the product claims against OpenAI's current documentation on 24 July 2026. The official Help Center confirms that Voice can start and coordinate tasks in Work and Codex on the ChatGPT desktop app for macOS and Windows, with paired iOS remote access. It also confirms one live Voice conversation at a time, plan and workspace-dependent availability, and separate usage treatment for Voice time and delegated agent work.
This rollout also changes an important detail from the original GPT-Live launch: Live initially did not support connected apps or plugins in ordinary Chat Voice. Voice inside Work or Codex now coordinates work through the tools and permissions of the selected experience. Those are related products, but they are not the same permission surface.
Link Map
| # | Resource | What it establishes |
|---|---|---|
| 01 | Riley Brown's Codex Voice test | Email triage, document creation, Excalidraw, parallel research, an iOS build, and paired iPhone control in one configured desktop environment. |
| 02 | OpenAI: Building with ChatGPT Voice | Official product demonstration of speaking with Work and Codex while tasks and agents continue. |
| 03 | ChatGPT Voice product page | The general Voice experience: start a live conversation, interrupt naturally, and review the transcript. |
| 04 | ChatGPT Voice Help Center guide | Current Voice options, availability, Work and Codex behavior, usage, data controls, paired iOS remote access, and limitations. |
| 05 | ChatGPT Work and Codex guide | The distinction between Chat, Work, and Codex; desktop availability; permissions; and the Voice setup path. |
| 06 | Official Codex use cases | Examples including inbox management, Computer Use, iOS debugging, background goals, QA, and connected workflows. |
| 07 | Download ChatGPT | The desktop app is the documented home for Voice in Work and Codex. |
| 08 | Riley Brown on YouTube and X | Creator credit and future follow-up tests. Riley explicitly describes this video as an early 30-to-45-minute experiment. |
What Actually Launched
The cleanest mental model is a three-layer stack:
| Layer | Job | What it should not be confused with |
|---|---|---|
| GPT-Live | Listens, speaks, handles interruptions, and maintains the live conversation. | It is not the repository, browser, plugin, or computer-control layer. |
| Work or Codex | Provides the task environment, project context, tools, permissions, agents, files, browser, and code workflows. | Work is for broader knowledge work; Codex remains the technical and software-development surface. |
| Voice control | Lets the person direct, inspect, prioritize, interrupt, and redirect work without typing every instruction. | It is not blanket permission for the agent to send, publish, buy, delete, or expose data. |
OpenAI documents Voice in Work and Codex on macOS and Windows. A paired iPhone can remotely access supported desktop Codex work, but standalone Voice in Work or Codex is not currently a mobile or web mode. The desktop remains the execution environment.
OpenAI's Official Demo
OpenAI's one-minute launch video is compact, but it shows the intended interaction pattern. A person asks Voice to draft launch copy, talks through an unfinished feature idea, asks Codex to inspect feature flags, brings in a new bug report, starts another thread, requests a pull request and root-cause analysis, and leaves a goal for the team to review later.
The important part is not the theatrical idea of talking to a computer. It is that the conversation remains available while work branches into separate tasks. Voice becomes a control plane over several visible work streams rather than a single answer box.
What Riley Tested
1. Email triage that stops at drafts
Riley asks the agent to find customer-support complaints, draft replies, and open the threads in the Codex browser. In his session, it opens 12 complaint threads and reports 11 saved drafts. This is a sensible first workflow because the agent does the reading and preparation while the replies remain reviewable.
The next instruction turns those issues into an anonymized Google Doc. This is where a Voice workflow becomes more useful than simple dictation: the spoken request triggers retrieval, synthesis, document creation, and a visible browser handoff.
2. Messaging with an explicit send decision
Riley asks the agent to make the document public and text the link to a contact. The demo shows the agent asking before sending the message, then asking again before sending a follow-up note. That is the right interaction shape, but teams should configure it rather than assume every connector and action will always pause in the same way.
3. A shared visual surface in Excalidraw
Riley opens an Excalidraw board in the in-app browser and asks Voice to inspect and extend it. The useful pattern is shared visual context: the person can speak loosely while the agent sees the current board, proposes additions, and makes a targeted change.
The demo also exposes a rough edge. A browser tab closes during session cleanup and has to be reopened. Voice does not eliminate the need to inspect what is actually on screen. It simply makes correction faster when the shared state drifts.
4. Parallel background research
Riley asks for an in-depth research note in Notion, then realizes he wants to keep talking. Voice moves the research into a separate task, keeps the live conversation open, reports progress, and later opens the finished note in the Codex browser.
This is arguably the most consequential part of the release. The conversation is no longer blocked by one long task. A live coordinator can dispatch work, keep a short control loop with the person, and return finished artifacts from background agents.
5. Building and revising an iOS app by voice
Riley asks for a Swift notes app, an in-app simulator run, visual references, and two additional parallel builds. He then uses Voice to inspect the simulator, close unrelated tabs, test the app, switch the design to light mode, and revise the bottom navigation.
The lesson is not that a spoken sentence reliably creates production software. The useful pattern is iterative direction: start a bounded build, ask the agent to test it, look at the rendered result, identify a specific change, and repeat. A voice interface makes that loop conversational, but code review, tests, security review, and platform checks still determine whether the build is ready.
6. Paired iPhone control
Riley pairs the ChatGPT iOS app with his Mac workspace, starts a separate support-summary task, asks for its chat name, and checks the state of the earlier app build. OpenAI's current documentation supports this paired iOS remote pattern.
Treat the phone as a remote control for desktop work, not as a second independent Codex environment. OpenAI does not currently document standalone Work or Codex Voice on mobile, and its current help page does not confirm Android remote support.
Documented Capability vs Creator Proof
| Claim | Status | Practical reading |
|---|---|---|
| Voice can start and coordinate Work or Codex tasks | Officially documented | Supported in the ChatGPT desktop app when available for the account and workspace. |
| Voice can coordinate multiple agents and background tasks | Officially documented and demonstrated | One spoken conversation can manage several task threads; only one live Voice conversation runs at a time. |
| Voice can use the selected experience's tools and permissions | Officially documented | Capabilities depend on configuration. Do not assume every connected app, local tool, or account is available. |
| Email, Docs, Notion, Excalidraw, and iOS simulator workflows | Demonstrated by Riley | These worked in his configured environment. Reproduce them with test data before granting production access. |
| Paired iPhone access to supported desktop Codex work | Officially documented and demonstrated | The phone steers desktop work. Standalone mobile Codex Voice is not documented. |
| Android remote support | Not confirmed in current official docs | Riley says it is coming, but builders should wait for an official availability notice. |
| Every consequential action will always request confirmation | Not established | The demo shows confirmation before messaging. Design explicit approval gates and test them for each connector. |
| Voice transcripts are an exact audit log | Officially contradicted | OpenAI says Voice transcripts may differ from what was spoken. Preserve tool logs and final artifacts separately. |
The Permission Model Is the Product
A voice interface encourages fast, informal instructions. That is its strength and its risk. People say "send that," "make it public," or "fix everything" more casually than they would complete a permission dialog. The system therefore needs more structure around consequential work, not less.
- Read: define which accounts, folders, repositories, and sites the task may inspect.
- Draft: allow reversible artifacts such as email drafts, documents, code branches, and proposed calendar events.
- Act: require confirmation for sending, publishing, inviting, purchasing, deleting, merging, or changing access.
- Verify: ask the agent to show the result, run tests, reopen the target, or report the final state.
- Log: preserve tool calls, changed files, recipients, links, approvals, and unresolved uncertainty.
Microphone, screen, audio-recording, accessibility, browser, local-file, and connector permissions should be granted only when the workflow needs them. Workspace administrators can also govern access to Work, local Codex, browser and network behavior, models, and defaults.
Copy-Ready Voice Agent Brief
Voice task brief
Outcome
- Complete: [one specific result].
Context
- Use: [project, folder, account, browser tabs, or connected apps].
- Ignore: [unrelated accounts, private folders, old projects].
Allowed actions
- Read, search, summarize, compare, draft, test, and open results for review.
- Start background tasks when work can safely run in parallel.
Approval required
- Before sending or publishing anything.
- Before changing sharing, permissions, subscriptions, or account settings.
- Before purchases, deletion, merging, deployment, or external submission.
Verification
- Show the final artifact or open it in a review tab.
- State what changed, what failed, and what remains uncertain.
- For code, run the relevant tests and show the result.
Definition of done
- [artifact exists]
- [acceptance check passes]
- [human approval is waiting at the stated boundary]Spoken prompts can be messy. The brief gives Voice a stable operating contract, so improvisation happens inside a defined boundary.
Five Strong Use Cases
| Use case | Why Voice helps | Required finish line |
|---|---|---|
| Feature planning | Ramble through the idea, invite questions, and dispatch research or prototypes without interrupting the conversation. | A written brief, assumptions, open questions, and a reviewable implementation plan. |
| Bug and incident triage | Describe symptoms while Codex inspects flags, logs, reports, and code in parallel. | Reproduction evidence, root-cause hypothesis, tests, rollback path, and a reviewed patch. |
| Inbox and meeting follow-up | Review priorities conversationally while the agent prepares summaries and drafts. | No-send drafts, recipient checks, sensitive-data review, and human approval. |
| Research orchestration | Keep the live discussion moving while separate tasks gather evidence and build documents. | Source links, dates, contradictions, confidence notes, and a final synthesis. |
| Hands-busy or accessibility workflows | Voice can reduce keyboard and mouse dependence for navigation, status checks, and direction. | Visible controls, mute and stop access, confirmation for consequences, and alternative input when speech is misheard. |
A Seven-Day Rollout
- Day 1: update and constrain. Confirm feature availability, choose one low-risk workflow, and list its required permissions.
- Day 2: establish a manual baseline. Complete the workflow yourself and record time, steps, expected result, and common mistakes.
- Day 3: use Voice in read-only mode. Ask it to inspect, explain, and open evidence without writing or sending.
- Day 4: add reversible drafts. Create one document, email draft, code branch, or proposed change and review it.
- Day 5: test interruption and recovery. Redirect a task, close a tab, deny a permission, introduce ambiguity, and confirm the agent stops or recovers visibly.
- Day 6: test the approval boundary. Attempt a send, sharing change, deletion, or publish action using test accounts. Confirm the configured human gate works.
- Day 7: compare accepted work. Measure completion time, corrections, tool cost, errors, and review effort. Expand only if the whole workflow improved.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Jarvis in Codex |
| 00:52 | Email triage demo |
| 02:20 | Summaries and texting |
| 04:44 | Excalidraw voice control |
| 06:32 | Background research tasks |
| 07:30 | Limits and safeguards |
| 10:31 | Build an iOS app by voice |
| 13:25 | Paired iPhone remote control |
| 15:36 | Wrap-up and next steps |
Bottom Line
The exciting part of this release is not that ChatGPT talks. It is that a live spoken conversation can now sit above Work and Codex, dispatch background tasks, coordinate agents, inspect progress, and return to a finished artifact without forcing the person to manage every intermediate step.
Riley Brown's "Jarvis" comparison captures the feeling. The responsible operating model is more grounded: one live controller, several bounded workers, narrow permissions, reversible outputs, visible verification, and a human decision before the system changes the outside world.
Start with research, drafts, and test environments. Once Voice can reliably understand the goal, preserve the evidence, recover from mistakes, and stop at the right boundary, it becomes a serious interface for work rather than a novelty microphone button.