I build Pageree, a remote MCP server that lets an assistant build a landing page, publish it after you approve a preview, and send the form's leads to your inbox or CRM. It works in Claude, Cursor and Codex, and since yesterday it's in ChatGPT's plugin directory. Getting there took four submissions, three rejections and almost two months.
This is what I'd tell anyone about to submit an MCP server to ChatGPT.
1. The submission is a project of its own
Before a reviewer looks at your tools, OpenAI wants:
- a verified identity for the publisher
- a challenge file on your MCP domain (
/.well-known/openai-apps-challenge) - a display name and a short description of 30 characters or less
- 5 positive and 3 negative test cases
- a demo video
- a demo account with sample data that a reviewer can use without signing up or 2FA
None of it is hard, but together it took me a couple of days. The demo account was the fiddliest: it has to have real-looking data in it, and nothing can stand between the reviewer and the first tool call.
2. Read the annotation rules twice
Every tool needs readOnlyHint, destructiveHint and openWorldHint set to explicit booleans. The guidelines are stricter than most MCP examples you'll find. Two lines are worth reading slowly:
"Set
readOnlyHint,destructiveHint, andopenWorldHintto explicit boolean values.""Being able to undo an action does not, by itself, justify setting
destructiveHinttofalse."
So a tool that overwrites or replaces anything (saving over a draft, publishing over a live URL, swapping where leads are delivered) is destructive, even if the user can roll it back. A tool that can reach an arbitrary destination (a webhook URL the user gives you, an image fetched from any URL) is open-world. ChatGPT asks for more confirmations as a result. That's the deal.
The hints are read from your live server, not from the package you upload. After you fix one, deploy, rescan the server in the dashboard, and only then submit again.
3. Record the demo before you trust your tool descriptions
My test cases said the user sees a preview and approves it before anything goes live. When I recorded the demo, ChatGPT saved the draft and published it in the same turn. My publish description already said not to do that. The model did it anyway.
What fixed it was a rule on the server. Publish now refuses a draft that was saved moments ago:
const savedAt = page.draft.lastUpdatedAt ?? page.draft.createdAt;
if (
options.reviewWindowMs &&
savedAt &&
Date.now() - new Date(savedAt).getTime() < options.reviewWindowMs
) {
throw new PublishRejectedError(
'Not published: this draft was saved moments ago, so the user has not ' +
'had a chance to review it. Give them the preview URL ' +
`${this.previewUrl(page.slug as string, page.draft.id)}, end your ` +
'turn, and call publish only after they approve this version in a ' +
'later message. Do not retry publish in this turn.',
);
}
The window is 30 seconds for an agent. The error message tells the model exactly what to do next, and it does it. If a test case depends on the user doing something first, enforce it on the server.
4. Freeze your tools while you're in review
Don't rename a tool or touch a description mid-review. I keep a snapshot of tools/list taken at submission and a small script that diffs the live server against it. OpenAI keeps scanning your server after approval too, so the habit is worth keeping.
5. Every round costs about two weeks
My last round took exactly two weeks, from resubmitting on 3 November to approval on 17 November. A rejection costs you that time again, so it pays to check everything against the guidelines before you press submit.
6. The listing can't sell
No prices, free trials or promotions in the listing copy. Tool descriptions can't push the model to pick you over other plugins, either. Write the copy like documentation.
Listed is not the same as found
It's been less than a day since approval and I see no traffic jump. I think being listed and being found are two different things. OpenAI's guidelines say plugins with "strong real-world utility and high user satisfaction may be eligible for … directory placement or proactive suggestions". So the listing is the starting line. People have to find it, use it and like it before ChatGPT suggests it on its own.
My plan is to send my own users through the listing: the setup docs now lead with it, and developer mode is only the fallback. I'll report back on what it does in a few weeks.
Has anyone here seen real traffic from ChatGPT's plugin directory? How long did it take to show up?
Pageree is at pageree.com, and the listing is here.
Top comments (0)