Skip to content

Getting an MCP Server Through ChatGPT Plugin Review: Two Rejections and What Fixed Them

OpenAI approved OrgPulse for ChatGPT after two sign-in rejections. Reviewer accounts, new-device codes and the credentials form explained, for MCP builders.

TL;DR

  • OpenAI approved OrgPulse for the ChatGPT plugin directory on 17 November 2026. I first submitted it on 8 October, so it took 40 days and three submissions. It is now listed in the ChatGPT plugin directory.
  • Both rejections gave the same reason: the reviewer could not complete sign-in. The OAuth setup was fine both times. The reviewer account was the problem.
  • The first rejection had two causes. I changed the reviewer password after submitting, and Clerk asked for an email code on a new device. A reviewer always signs in from a new device and cannot read your inbox.
  • The second rejection came after I emailed the new password to the review team. Reviewers only use the credentials saved in the portal's Review information form, and resubmitting does not ask you to update them.
  • OpenAI's plugin scan also flagged a tool annotation. A tool that sends alerts to Slack and email has to say openWorldHint: true.

What You'll Learn

  • What "unable to complete your sign-in or OAuth flow" usually means, and how to tell whether the reviewer got in at all
  • Where reviewer credentials live in the submission portal, and why replying to the rejection email does not update them
  • How to set up a reviewer account that works on the first try: no MFA, no new-device codes, a password you do not touch
  • The tool annotations the automated scans check, and how I justified openWorldHint for a tool that never touches Salesforce
  • How to write positive and negative test cases against a demo org whose data does not drift between review windows

The Problem

In my post on connecting ChatGPT and Claude to Salesforce I said OrgPulse was waiting for approval in the ChatGPT directory. That was on 9 November. By then it had already been rejected twice, and I had not worked out the second rejection.

OrgPulse is a remote MCP server. ChatGPT signs in to it with OAuth through Clerk, and OrgPulse then reads the customer's Salesforce org as a read-only integration user. I had tested that flow many times in ChatGPT developer mode. It worked in a fresh incognito window. And OpenAI still came back twice with the same sentence: they could not complete the sign-in or OAuth flow, and the credentials must need "no additional setup or verification".

That sentence covers several different failures: a wrong password, an expired password, an MFA prompt, an email code, a CAPTCHA, or an OAuth discovery error. The rejection email does not say which. If your OAuth works in testing, the first thing to suspect is the reviewer account, not your code.

Common questions this article answers:

  • Why was my ChatGPT plugin rejected for sign-in when OAuth works for me?
  • How do I update reviewer credentials after a rejection?
  • What does OpenAI's review need from a demo account?
  • Which MCP tool annotations does the review check?

Quick Answer

Give the reviewer a dedicated account that signs in with email and password and nothing else. Turn off MFA, email codes on new devices, magic links and bot challenges for that account. Put its credentials in the portal under Review information, not in an email. Do not change the password while a review is open. Before each resubmission, revise the submission, open Review information, check the reviewer credentials step and save the current password there. Then run the whole reviewer path in a fresh incognito window on the day you submit, and check that the reviewer inbox received no verification email.

The timeline

All dates are New Zealand time.

Date What happened
8 October First submission received
12 October I changed the reviewer account's password, three days into the review
24 October Rejected: unable to complete sign-in or OAuth flow
28 October Fixed the account, resubmitted, and emailed the new password in reply to the rejection
29 October Rejected again, same reason
3 November Saved the new password in the Review information form and resubmitted. Emailed to ask what exactly failed
7 November Review team: save updated reviewer credentials under Review information; the account must work without MFA or other verification
12 November Review team: the earlier attempt showed "Password is incorrect"
17 November Approved, version 1.0.0, and published to the directory

Three submissions, all of the same version 1.0.0. Nothing in the server changed between the second and third.

Rejection one: the reviewer account, twice over

Two separate things broke the first review, and either would have been enough.

The first was my own doing. I reset the reviewer account's password three days after submitting, and the form still had the old one. Nothing warns you that the credentials you entered no longer work.

The second was Clerk, which OrgPulse uses as its OAuth authorisation server. When a password sign-in comes from a device it has not seen, Clerk sends an email code to finish the sign-in. That is a sensible default for real users. A reviewer, though, is always a new device and cannot read the inbox, so for them the sign-in can never finish.

The reviewer account's own inbox shows both. On the morning of 12 October it received a verification code, then a "your password has been changed" email two minutes later. After the second rejection, Clerk showed no session at all for the reviewer account. So the reviewer never got past the password, and it was not an OAuth problem.

The fix was to turn off new-device verification in the production Clerk instance, set a new password, and leave it alone. If your provider has a bot challenge or CAPTCHA on sign-in, check that too.

Rejection two: the password was in the wrong place

On 28 October I resubmitted and replied to the rejection email with the new password, because the resubmit flow had no field for credentials. The next day OrgPulse was rejected again for the same reason.

Reviewers sign in with the credentials saved in the portal form, and only those. A password in an email thread is not used. The form can still be edited, but getting to it is not obvious. You press Revise submission, open the draft's menu, choose Review information and go to the reviewer credentials step. Saving a new submission does not ask you to look there.

I worked that out and resubmitted on 3 November with the password in the form. I also asked the review team what exactly they had seen. The answer came on 12 November: "Password is incorrect." It was the most useful thing anyone told me during the whole review, and it arrived after I had already fixed it. If you get the generic rejection, ask for the exact error anyway, but do not wait for the reply before you check the form.

One more consequence: a password you email to a review team now sits in that thread. Rotate it once the review closes.

How I knew the third review got in

On 17 November the reviewer account received OrgPulse's "new device signed in" email. Ten minutes later the approval email arrived. Between 12 October and that morning, the reviewer inbox received no sign-in notification at all. If your app sends sign-in notifications, point them at an inbox you watch during review. They are the only direct sign you will get that the reviewer reached your app.

What the automated scan checked first

OpenAI also runs a scan over the tools your server declares. It flagged one thing on OrgPulse: configure_watch declared openWorldHint: false.

My reasoning had been that the tool never touches Salesforce. It only saves an alert rule in OrgPulse's own database. But a saved watch later sends alerts to whatever Slack webhooks and email addresses it names, and that is an external system, so the hint has to be true. The other eight tools read the user's own org or OrgPulse's own data and stay false.

This is how the nine tools ended up:

Tool readOnlyHint destructiveHint openWorldHint
Eight read tools (expiry, limits, storage, licences, two access explainers, two alert readers) true false false
configure_watch false true true

configure_watch is destructive because replacing a watch overwrites its earlier thresholds and channels. All nine are idempotent. I wrote a one-sentence justification for each hint for the review notes. That forced me to check claims like "reads only", which I would otherwise have taken on trust.

Claude's directory, which I submitted to at about the same time, flagged something different: every tool was missing annotations.title. The fix was one line in the registration loop to copy each tool's title into its annotations, plus a test. The two directories check different things, so run each one's checks against your server before you submit.

The demo org has to stay true

The review form wants test prompts and the tool calls you expect, including negative cases. Mine run against a Salesforce Developer Edition org with synthetic data only, connected to a workspace that belongs to the reviewer account. I gave the reviewer account a complimentary paid plan so the alerting tools work. On the free plan those tools return a short message with a link to the pricing page.

The positive cases ask what expires in the next 30 and 60 days, how close the org is to its API limit, and why one user can see Account.Rating and another cannot. I also wrote three negative cases:

  • "Deactivate Bob Synthetic and delete the permission set assignment that expires in 10 days." No tool writes, so ChatGPT should say OrgPulse is read-only.
  • "What expires in my org Globex Synthetic?" That org is not connected, so the tool returns an error naming the orgs that are, and no data from any other org.
  • "Write me a poem about autumn." No OrgPulse tool should be called at all.

Expiry data goes stale, though. My expected answers mention permission set assignments ending in 10, 25 and 45 days. A Developer Edition org does not expire on a timer, but those dates do, so a review that starts two weeks late fails your own test cases. I re-seed the demo org before every review window.

Re-seeding turned up a Salesforce quirk. Once a time-limited permission set assignment has expired, plain SOQL no longer returns it, yet it still blocks a new assignment of the same permission set with DUPLICATE_VALUE. My seed script looked for existing assignments, found none, and then failed on insert. Querying with --all-rows finds the expired rows, so the script can re-date them instead.

Things that did not cause trouble

None of these came up in either rejection. They are all in OpenAI's submission guidelines, so get them right before the first submission:

  • Domain verification. The app domain serves a token at /.well-known/openai-apps-challenge, as the bare token rather than JSON.
  • OAuth discovery. ChatGPT found Clerk through the protected resource metadata and registered itself with a Client ID Metadata Document, as described in the earlier post.
  • No commerce inside ChatGPT. When a free user asks for a paid feature, the tool answers with text and one link to an informational pricing page, never a checkout. The long description has no upgrade pitch.
  • No widgets. Every tool answers with text and structured content, so there was no content security policy to declare.
  • The privacy policy. The guidelines expect it to list the categories of personal data, why you collect them, who receives them and how long you keep them.

Frequently Asked Questions

Q: What does "unable to complete your sign-in or OAuth flow" mean in a ChatGPT plugin rejection?

A: The reviewer could not get from Connect to a working, signed-in plugin. That covers a wrong or changed password, an MFA prompt, an email code, a CAPTCHA, and genuine OAuth discovery errors. If the flow works for you in a fresh incognito window, look at the reviewer account first. Check your authorisation server for a session from the reviewer account. If there is none, they never got past the password.

Q: How do I update reviewer credentials after a ChatGPT plugin rejection?

A: Edit them in the portal. Press Revise submission, open the draft's menu, choose Review information and update the reviewer credentials step before you submit again. Replying to the rejection email with a new password does not change what the reviewer uses.

Q: Can the reviewer account use MFA?

A: No. OpenAI asks for credentials that need no additional setup or verification. That rules out MFA, and also email codes on a new device, magic links and bot challenges. A reviewer is always on a new device, so check your identity provider's new-device setting in particular.

Q: How long does ChatGPT plugin review take?

A: For OrgPulse, 16 days to the first decision and 40 days from the first submission to approval, with two rejections in between. The second review came back the next day and the third took 14 days. That is one data point, so treat it as an example, not a benchmark.

Q: Does a tool that only writes to my own database need openWorldHint true?

A: It depends on what happens afterwards. If saving the record later causes messages to go to Slack webhooks or email addresses the user supplied, OpenAI's scan treats the tool as touching an external system. OrgPulse's configure_watch was flagged for exactly that and now declares openWorldHint: true.

Q: Do I need a demo org for a Salesforce MCP server review?

A: You need data the reviewer can query straight after signing in. For OrgPulse that meant a Developer Edition org with synthetic users and time-limited permission set assignments, already connected to the reviewer's workspace. Re-seed it before each review window, so dates in your expected answers still hold.

Key Takeaways

  • A sign-in rejection usually means the reviewer account failed, so check the account before you debug OAuth.
  • Give the reviewer email and password only: no MFA, no new-device codes, no CAPTCHA.
  • Reviewers use the credentials saved in the Review information form. Update that form on every resubmission, and never change the password mid-review.
  • Run the reviewer path in a fresh incognito window on submission day, and watch the reviewer inbox for verification emails.
  • Declare openWorldHint: true on any tool whose effects reach Slack, email or another outside system, even later.
  • Seed demo data so your expected answers are still true on the day someone tests them.

What's Next?

Recommended reading:

Action items:

  1. If you have a plugin or connector in review, sign in as the reviewer account in an incognito window today and confirm no code or prompt appears.
  2. Open Review information on your submission and check the saved password is the current one.
  3. Read every tool's annotations against everything the tool causes, including what happens after the call returns.
  4. To try OrgPulse, open its ChatGPT plugin listing and choose Install plugin, or add it to Claude as a custom connector. The free plan checks one org on demand at orgpulse.cloudcounsel.co.nz.

Resources & References

Responses

Checking your session.

Loading responses.