Most guides on ChatGPT agent marketing still describe agent mode: type /agent, watch it click. OpenAI’s help page now says ChatGPT agent is no longer available and points to ChatGPT Work. This page is for founders who pay for ChatGPT, do not use a terminal agent yet and want to launch a SaaS on directories. Every OpenAI fact below comes from OpenAI’s own pages, with the date we checked it.
What happened to ChatGPT agent mode
We read OpenAI’s pages on September 28, 2026. OpenAI changes these products often, so open the linked source before you rely on a detail.
- Agent mode status
- The help article on ChatGPT agent now opens with: “ChatGPT agent is no longer available.” It points to ChatGPT Work and the cloud browser. We did not find a published retirement date.help.openai.com ↗ · Checked on September 28, 2026
- What agent mode was
- A mode with its own virtual computer: a visual browser, a text browser, a terminal and connectors. It asked before actions with real-world consequences, such as a purchase.openai.com ↗ · Checked on September 28, 2026
- What replaced it
- ChatGPT Work, introduced on July 9, 2026: an agent for longer tasks that researches, uses connected apps and files, and delivers documents, spreadsheets and slides. Paid plans only, not Free or Go. The Atlas browser stopped working on August 9, 2026.help.openai.com ↗ · Checked on September 28, 2026
- Cloud browser
- Work on the web and on mobile uses its own browser on a separate computer, without your tabs, cookies, saved passwords or sign-ins. It keeps running after you leave and pauses for input, sign-in or confirmation.help.openai.com ↗ · Checked on September 28, 2026
- Sign-ins
- Since August 25, 2026, Work can finish tasks on some signed-in sites on Plus and Pro. You type credentials and 2FA codes into a secure form the model cannot see.help.openai.com ↗ · Checked on September 28, 2026
- Blocked websites
- Some sites restrict automated browser agents, so a task can fail even if the site works in your own browser. Site owners can allowlist the cloud browser, which signs its requests.help.openai.com ↗ · Checked on September 28, 2026
- Your own browser
- In the desktop app, a ChatGPT browser extension lets Work and Codex act in Chrome, Edge, Brave, Opera or Vivaldi on sites where you are already signed in. By default it asks before using each new website.learn.chatgpt.com ↗ · Checked on September 28, 2026
- Local files
- Work on the web and on mobile cannot open files on your computer. The desktop app can, for the folders you allow.help.openai.com ↗ · Checked on September 28, 2026
- Codex
- OpenAI’s coding agent is included in every ChatGPT plan, Free and Go too, with usage limits that vary by plan. The Codex CLI signs in with your ChatGPT account.help.openai.com ↗ · Checked on September 28, 2026
- MCP in Codex
- Codex can start local MCP servers that you add with codex mcp add or in ~/.codex/config.toml.developers.openai.com ↗ · Checked on September 28, 2026
Where ChatGPT Work needs you for launch submissions
Directories are built to keep unattended bots out, and OpenAI says so openly. Plan for these stops:
- A second set of sign-ins
- The cloud browser does not share your Chrome’s sessions. You sign in to every directory again through the secure form, and some sign-in flows are not supported there.
- Blocks and bot checks
- Cloudflare, reCAPTCHA and similar checks can block the cloud browser or stop it mid-form. ChatGPT then asks you to take over. LaunchRepo’s playbooks never try to get around a challenge.
- Email verification
- Most directories confirm by email. Work only sees your inbox if you connect Gmail. Clicking the link yourself is safer: a mailbox plus an agent is what prompt injection aims at.
- Paid upsells
- ChatGPT asks before a payment, but a paid fast track can look like a normal form step. Say in the task that it may not buy anything.
- Usage allowance
- Work draws on your plan’s shared Work and Codex allowance. OpenAI says use varies by task and publishes no number per submission.
- Knowing what “live” means
- A thank-you page is not a listing. Some platforms queue for weeks, some review by hand. A chat log is a poor tracker of what went live.
ChatGPT agent mode for marketing today: which part to use
ChatGPT Work in the cloud is good at the thinking part of a launch: researching where similar products listed, drafting descriptions and preparing one form while you watch. It keeps working while you are away. That suits research more than submissions, which end at a sign-in, a check or an email.
In the desktop app, ChatGPT can act in your own signed-in browser through its extension. That is closer to what a launch needs, but LaunchRepo’s browser guide supports one connection only, Chrome DevTools MCP, and asks agents not to use extension-based control. So we do not call that route supported.
Codex is the part of your ChatGPT plan that fits LaunchRepo: it reads a cloned folder, runs Python and can use Chrome DevTools MCP with your everyday Chrome. LaunchRepo is not a GPT or a plugin. It is a private GitHub repository with a runbook, 299 directory playbooks, a product template and a tracker.
Running LaunchRepo with your ChatGPT plan
LaunchRepo asks three things of an agent: read the files, run Python 3.10+, and drive a Chrome where you are signed in, through Chrome DevTools MCP. Work in the cloud cannot reach your folder, your Python or your Chrome. The Codex CLI can, and it signs in with the ChatGPT account you already pay for.
- Install the Codex CLI and sign in
- Install it with OpenAI’s installer, npm or Homebrew and choose “Sign in with ChatGPT”. Your plan’s Codex usage applies; no API key is needed.
- Clone the repository and start Codex there
- After purchase you accept a GitHub invitation, clone the private repository and run codex inside the folder. It reads AGENTS.md first.
- Add Chrome DevTools MCP
- Register it with codex mcp add or in ~/.codex/config.toml, switch on remote debugging in chrome://inspect (Chrome 144 or newer) and approve the prompt. Then check that the tab you can see actually moves.
- Keep ChatGPT for research and copy
- Use Work to research and draft descriptions, check the facts, then put the approved text into product.json. Codex fills every form from that file.
On the Codex side the browser server is two commands, run once in any terminal. Everything after the double dash is the server line from LaunchRepo’s browser guide:
codex mcp add chrome-devtools -- npx -y chrome-devtools-mcp@latest --autoConnect --no-usage-statistics --no-performance-crux
codex mcp listFormat as documented by developers.openai.com, checked on September 30, 2026. ↗
codex mcp list should now show chrome-devtools. Switch on remote debugging in Chrome, let Codex open example.com in a blank tab, allow the prompt and confirm the tab in front of you moved.
Connect your coding agent to your Chrome →
In Codex, paste the public start prompt from START.md, unchanged. Its first line recommends Claude Code CLI or Codex CLI, so with Codex you are on the default path. Do not paste it into ChatGPT Work: Work on the web cannot open the cloned folder on your computer.
Recommend Claude Code CLI or Codex CLI in a terminal instead of the desktop apps for best results. If session metadata confirms desktop hosting, explicitly advise me to switch to the CLI and resume this same private workspace. Do not infer the host from a CLI engine version alone.
Read AGENTS.md, then follow "Start the dashboard for the owner". Create my private
workspace if I do not have one yet, start the dashboard, and give me the link.
Then lead ONBOARDING.md one phase at a time. Reuse recorded permissions instead of
asking again unless their scope is missing or conflicting. Actually test my visible
existing Chrome profile and intended identity as BROWSER-SETUP.md requires; use the
OS fallback if needed. Use one platform/browser worker by default. Build the normal
run only from suitable free-ready routes, keep research separate, and preserve any
draft that needs one exact human CAPTCHA/2FA action. Never claim browser control
bypasses or reduces those challenges.The playbooks and instructions are English Markdown. Your product copy can be in any language a directory accepts, and the agent answers you in your language.
When ChatGPT or Codex gets stuck
Three situations that come up when a launch moves from ChatGPT to Codex, with the way out OpenAI’s docs describe.
- Work stops on a site
- Some sites restrict automated agents in the cloud browser even though they work in your own. Move that platform to Codex and your Chrome, or submit it by hand with its playbook open.
- Work cannot find the repository
- Work on the web cannot open files on your computer. Run the start prompt in Codex inside the cloned folder instead; the desktop app only reaches the folders you allow.
- Codex does not list the server
- Run codex mcp list. If chrome-devtools is missing, add it again; a project-level .codex/config.toml only counts in projects you trust, while ~/.codex/config.toml always does.
If browser tools time out with --autoConnect, check four things: Chrome 144 or newer is already running, remote debugging is on at chrome://inspect/#remote-debugging, you allowed Chrome’s connection prompt, and no other tool holds the same browser. Chrome DevTools MCP loads every open tab, so close the ones you do not need.
ChatGPT Work or Codex: who does what
One launch, split three ways: Work in the cloud, Codex with LaunchRepo on your machine, and you.
| Task | ChatGPT Work (cloud) | Codex + LaunchRepo | You |
|---|---|---|---|
| Launch copy | Drafts and refines it | Uses the approved text in product.json | Check the facts once |
| Platform research | Researches the web, unverified | Starts from 299 dated playbooks | Pick this week’s batch |
| Signed-in submissions | Cloud browser with its own sign-ins; some sites block it | Your own Chrome through DevTools MCP | Sign in once per site |
| Challenges and email links | Asks you to take over | Stops that platform, notes a human handoff | Handle it in your Chrome |
| What went live | Only in the chat history | submissions.json plus the dashboard | Open the public URL and confirm |
Common mistakes
- Hunting for the agent mode toggle
- It is gone. Choose Work, or Codex. Do not upgrade your plan hoping the old button comes back.
- Pasting passwords into the chat
- OpenAI says never to. Use the secure sign-in form or take over the browser, once per platform.
- Setting website access to “Always allow”
- OpenAI marks it as not recommended. Directory pages are written by strangers, and any of them can carry a prompt injection.
- One Work task for the whole list
- A single long task mixes sign-ins, blocks and mail links on dozens of sites. Work in small batches with a duplicate check first, and track each platform on its own.
Questions
Is ChatGPT agent mode gone?
Yes. OpenAI’s help article says ChatGPT agent is no longer available and sends you to ChatGPT Work, which launched on July 9, 2026, and its cloud browser. We found no published retirement date.
How do I use ChatGPT for launch marketing now?
Choose Work and describe the outcome: the product, the websites, what it may and may not do. Approve website access, sign in through the secure form and review every confirmation. For repeatable directory work, run Codex with a playbook repository.
Can ChatGPT Work submit my SaaS to directories?
Some, yes, after you sign in. Sites that block automated agents stop it, and it does not know each platform’s rules or track what went live.
Does LaunchRepo work with ChatGPT?
Through Codex, yes. Codex is included in ChatGPT plans, and the CLI is one of the two routes START.md recommends. LaunchRepo does not run as a GPT, a plugin or a Work task.
Can ChatGPT solve CAPTCHAs on submit forms?
OpenAI does not promise that. Its docs say sites may block agents and ChatGPT asks you to take over when stuck. LaunchRepo records a challenge as a human handoff.
What does it cost?
Your ChatGPT plan covers Work and Codex up to its allowance. LaunchRepo costs €119.00 once at the current price, for unlimited own products and any agent.
Give your agent the launch playbooks
The current price is €119.00, paid once, for unlimited own products and any agent. The price does not rise.
See pricing