Voice is becoming a real interface for getting work done in ChatGPT and Codex — not just for dictating a message, but for discussing tasks, launching them, and following up on results. This guide walks through choosing the right voice mode, setting it up, and running through a first end-to-end scenario, from a spoken instruction to a verified result.
It draws on a widely shared thread that broke down these workflows in detail, cross-checked against OpenAI's documentation. The scenario involving a second bot is a personal workflow example; the rest are adapted from common use cases.
Which voice mode should you use
| Mode | Best for | What to keep in mind |
|---|---|---|
| Dictation | Speaking a request, editing it, and sending it | This is single-message input only |
| Voice in ChatGPT (chat) | Discussing an idea, asking questions, rehearsing a meeting | Capabilities depend on which Voice variant you have |
| Voice in the desktop app, with Work or Codex | Assigning work, checking on tasks, passing along clarifications | Requires the task to have access to the right files and tools |
In conversational ChatGPT, the Live variant supports natural dialogue, web search, memory, text, and images. Video, screen sharing, connected apps, and plugins aren't yet supported in Live. For video and screen sharing on iOS and Android, an Advanced variant is available if your account has access.
In the desktop app, voice lets you coordinate work across Work and Codex tasks. You can check on the status of a job, pass notes to another task, and switch back. This requires that the task support voice and run on a connected machine.
Work and Codex share the same underlying tool-use capabilities. Work is built around documents, research, and everyday tasks; Codex lets you continue similar work in a familiar development environment. For tasks involving your own local apps and files, choose local execution.
Turning on voice and app access
To get started, you need the desktop app with Voice available, a microphone, and a specific task. Voice is available on supported paid plans, though availability also depends on rollout and organization settings.
- Open a task and click Start voice chat (or Start new voice chat for a new conversation). On first use, allow microphone access and pick a voice.
- On Mac, to discuss an open window, enable Settings → Voice → Screen context. Once you ask the assistant to look at your screen, it captures a snapshot and any visible text from the active window.
- If you need it to click buttons and enter data in other apps, install Plugins → Computer Use, turn on its server and skill. On Mac, grant screen recording and Accessibility permissions. Access to specific apps is checked separately.
- Connect your sources — the relevant folder, email, documents, or browser. Use a dedicated plugin where one exists; Computer Use is useful for interface-driven actions.
Screen context passes along window contents; controlling that window is handled by Computer Use. Visible text may include content beyond what's currently on screen. Make sure the right window is open before asking for a review.
Your first task and checking the result
Start with a small document in an accessible folder. Name the task "Meeting Prep" and say:
Read the meeting notes file. Highlight the open questions we still don't have answers for, and draft a short agenda. Note which file each point comes from. Save the result as a separate file next to the original.
Once work begins, follow up by voice:
Add, for each question, what decision we need to make. At the end, tell me where the finished file is.
Open the result and check three things: the right source file was used, your last clarification was incorporated, and the file was actually saved. This confirms the whole loop — speech, execution, follow-up, and result.
For future tasks, use one simple structure: where to work → what to do → what to return → where to stop. For example: "In the connected inbox, find this week's thread with the client, pull together the agreed points with links, draft a reply, and stop before sending."
Useful scenarios
A good habit is to keep a voice conversation open and switch on the mic as soon as a task comes up. Here are twelve practical scenarios grouped by theme — pick the ones that match the sources and tools you already have connected.
Managing ongoing work
Find out where your input is needed. With several tasks open: "Look at my current tasks. Tell me which ones are done and which are waiting on me. Open the first task that has a question for me." The result is a specific task and a clear reason it's stalled; inaccessible tasks won't show up in the summary.
Pass a note to another task. "Switch the conversation to the 'Client Website' task. Ask it to simplify the signup form down to just name and phone number. Then bring me back to 'Meeting Prep'." Check that the note actually landed in the right task — using the exact task name reduces the risk of mixing things up.
Wrap up your day. "Summarize today's tasks: finished results, work still in progress, and any decisions still waiting on me. Draft a note for tomorrow with links." The result is a saved note; check long-running tasks by their most recent update.
Turning a conversation into a finished document
Work through a rough idea. "I'm going to explain an idea. Ask me one question at a time until it's clear who this is for and what the first version needs to do. Put together a brief first, then we'll talk about starting the build." The output is a brief you can review before development starts.
Prep for a meeting. With notes and email connected: "Pull together what's changed with this client since our last meeting, and any promises I haven't followed up on. Then play the client and ask me tough questions." Check the source summary first — the rehearsal questions will be the model's guesses.
Turn a long story into structure. "I'm going to walk you through what happened with this client. When I say 'done,' put together a timeline, a list of open questions, and a handoff note." The result is three sections in one document. Before handing it off, double-check names and dates it picked up by ear.
Working with files, email, and numbers
Sort through email. "In today's inbox, find questions I haven't answered, promises I made, and anything due tomorrow. Draft replies for the urgent ones." The result is a list with source emails plus drafts — treat sending as a separate step after review.
Track down a lost file. "In the project folder, find the proposal with three pricing options — we discussed it last month. Show me the candidates and compare the two latest versions." Check file paths and dates: a similar name doesn't guarantee it's the right document.
Recalculate a budget. "In a separate copy of the budget, move the hiring date from September to November. Show me which months and totals changed." The result is a separate scenario plus a diff against the original. To verify, ask for a list of the assumptions and formulas that changed.
Hand off an independent check. "Create a separate task: reconcile this report against the source spreadsheet. Return discrepancies in the totals and findings without assuming the data is correct." The checking task needs both files. Its output should be a list of specific discrepancies tied to cells or excerpts, not a general quality judgment.
Using the screen and browser
Show an interface bug. On Mac with Screen context enabled: "Look at this open window — the menu is covering the payment button. File a task, reproduce the issue, and prepare a fix." The screenshot helps explain the bug; fixing it requires project access. Verify by repeating the same steps after the change.
Fill out a form. With a browser, resume, and portfolio available: "Fill out this application using my files. Ask me about anything missing. Stop before submitting." The result is a completed form you review before confirming.
My workflow: Codex → a second bot → Codex
I use Codex's voice mode alongside control of other apps on my machine — including instructing it to reach out to an agent running in another bot, pull context from it, and bring that back. This is how I gather information from X (formerly Twitter).
The sequence looks like this: I give a spoken instruction to Codex → Codex opens the other app → hands the task to a specific agent → gets a response → brings that context back into the original Codex task and keeps working.
For this kind of handoff, it helps to name the agent and the expected format up front:
Open [the other app]. Hand the 'Researcher' agent the task of gathering posts on this topic from X. Get back a response with links to the original posts and their dates. Bring what you found into the current Codex task, separate facts from the agent's interpretation, and continue drafting. If any sources are unavailable, say so.
"Researcher" here is just an example name — use whatever your own agent is called. This requires the other app to be running, its interface to be accessible, and Codex to have permission to interact with it. This is a personal workflow built on available tools, not a built-in integration officially offered by OpenAI.
Always check the returned links and the completeness of the results. If the task was to collect someone's posts, ask for a count of what was found and whether pagination was completed. A secondhand summary from another agent, without the original posts, isn't enough to verify the material.
Computer Use has real limits — official documentation excludes automating terminal applications and ChatGPT itself. So a workflow involving another agent depends heavily on which interface you're using to reach it.
Limits and troubleshooting
Discussions around higher-tier plans, cheaper model choices, and voice usage don't amount to a promise of unlimited voice access across all modes — treat any such claims with caution.
According to documentation, Voice in the desktop app draws from Codex's overall usage budget: voice itself is billed at roughly $0.05 per minute, while the model's actual work on a task is billed separately based on tokens. For credit-based plans, voice usage runs around 1.25 credits per minute. Always check current pricing pages for up-to-date terms. Conversational Voice in ChatGPT has its own limits depending on your plan.
Muting the microphone is not the same as ending a voice session. Don't assume a muted session is free — when you're done talking, click Stop voice chat.
| Issue | What to check | Sign it's fixed |
|---|---|---|
| No voice button on an older task | App and machine updates where the task runs; feature availability for your account | The button appears, or a new voice conversation becomes available |
| Voice won't start | Microphone access and any other active voice session | Conversation starts; only one desktop session can be active at a time |
| Agent can't see the app | Computer Use plugin, system permissions, and app access | Agent correctly describes the window in question |
| Mouse gets hijacked on Windows | Computer Use runs in the active Windows session | You pause your own input while the task runs |
| Response is just a promise | Ask for a file, link, or visible change | Result can actually be opened and checked |
Start with one task and one source. Once a voice instruction reliably produces a verifiable result, start layering in task-switching and cross-agent workflows.
A good place to begin is with a single recurring task — sorting email, prepping for a meeting, or gathering sources. It's easier to set up and validate against your own real data before expanding to more complex, multi-agent workflows.