You’re mid-coding session. You say, “Hey ChatGPT, look at this Python function, find the bug, and suggest a fix.” Before it finishes responding, you add, “And draft a quick email to the team saying I’ve found the fix.” The system handles both without you touching the keyboard. That’s not a demo reel. It’s what GPT-Live’s full-duplex voice makes possible right now in the ChatGPT desktop app. It works differently from every voice assistant you’ve used before, because it’s listening, thinking, and speaking at the same time.
What Actually Changed: Full-Duplex
This isn’t just another push-to-talk feature. Previous voice assistants, including the old ChatGPT Voice, are turn-based. You speak, you wait. It speaks, it waits. GPT-Live is full-duplex, which means it can process your input while it’s generating its own output. You can interrupt it, correct it, and stack commands naturally. This is the key that unlocks agentic control.
The stack you’re controlling is specific:
- ChatGPT Desktop App: The only place this works (macOS and Windows).
- ChatGPT Voice: The user-facing feature.
- GPT-Live: The full-duplex engine running it all.
- Codex: The agent for code generation and dev tasks.
- ChatGPT Work: The agent for general productivity, file analysis, and content creation.
You switch between agents with explicit commands like “Hey ChatGPT, open Codex…” or “Switch to Work…”. This is a real change in how you interact with the system. It’s available on Plus, Pro, Business, Edu, and Enterprise plans. There is no developer API at launch.
Setup: Hands-Free in Under 5 Minutes
Getting this running is fast.
- Confirm your plan and update to the latest version of the ChatGPT desktop app.
- Activate Voice Mode by clicking the headphone icon in the bottom-right of the chat window.
- Grant microphone permissions when prompted.
That’s it. Now test it with two distinct agent commands.
First, a Codex task:
“Hey ChatGPT, open Codex and write a simple Python script using Flask to create a web server that returns ‘Hello, World!’”
Codex will speak a confirmation and drop the app.py file directly onto your desktop.
Next, switch to the Work agent:
“Switch to Work. Read the file ‘project_brief.txt’ on my desktop and give me a three-bullet summary.”
ChatGPT Work will read the local file, speak the summary aloud, and display the text in the app window.
Tips for Success
If it misfires, don’t just repeat the command louder. Rephrase it. Check your system’s microphone permissions if it seems unresponsive. Be specific with agent names. The model needs the explicit “open Codex” or “switch to Work” to route the task correctly.
Workflow 1: The Agentic Coder (Codex)
Let’s run a real dev task, complete with a predictable failure. Before we start, know that Codex interacts with your local file system directly, not in a sandbox. It will ask for permissions to read/write to specific directories like your desktop or project folders. Be mindful of what you grant it access to.
First, create a project folder on your desktop named stock-ticker.
Then, give the first command:
“Hey ChatGPT, open Codex. In the ‘stock-ticker’ folder on my desktop, create a Python file called ‘ticker.py’ that uses the yfinance library to get the current stock price for Apple and prints it to the console.”
Codex will confirm and create the file. Now, try to run it.
“Run the script.”
This is where it breaks. Codex responds that the script failed because the yfinance library isn’t installed in your current Python environment. It then correctly generates the command you need: pip install yfinance. It can’t run the install for you. You have to copy that command and run it in your terminal manually. This is a key friction point. After you install the dependency, you can issue the run command again.
“Run the script again.”
This time it works. The agent speaks the current AAPL price aloud. It’s a useful workflow, but the dependency management is still on you.
Workflow 2: The Content Assistant (ChatGPT Work)
Now for a business task. Put a simple CSV file on your desktop named q3_sales.csv with columns for Region and Sales.
Start the chain:
“Hey ChatGPT, switch to Work. Analyze the ‘q3_sales.csv’ file on my desktop and tell me the total sales for each region.”
The agent will read the file and speak the results, for example: “Total sales for North America were $150,000, and Europe was $110,000.”
Next, generate content from that analysis:
“Draft a short report with the headline ‘Q3 Regional Sales Performance’ using that data.”
The app displays the drafted report in the window and confirms with a spoken message. Now, let’s switch contexts.
“Okay, now draft a Slack message for the #sales channel summarizing the top performing region in one sentence.”
Here, the system shows how it manages state. It will say something like, “Got it. The report is saved as a draft. Here is the Slack message,” and then it generates the new text. It doesn’t lose the report, but it requires that handoff. It’s not a seamless multi-document juggle; it’s a serialized task handler.
The Missing API
For builders, the first question is always “Can I integrate this?” Right now, the answer is no. There is no developer API for GPT-Live.
This means you can’t trigger these voice-driven agents from your own scripts, a CI/CD pipeline, or internal tools. You can’t build custom voice commands into your IDE or orchestrate a sequence of agent actions programmatically. It’s a closed loop inside the ChatGPT desktop app.
Related
- Claude Code Ships Agentic Loops: How Anthropic’s CLI Became the Most Dangerous Tool in Your Terminal
- Build with Antigravity: Run Parallel AI Agents in Google’s Free VS Code Fork
An API would change this completely. Imagine voice-triggered deployments, headless agents running on servers activated by a voice command, or deep integration with tools like Linear or Jira. That’s the potential. For now, what shipped is a genuinely useful, self-contained tool for hands-free interaction on your local machine. It’s a solid first step, not the whole platform.