I’m posting this in case it helps OpenAI or other users troubleshooting long-running ChatGPT plugin connections.
What happened
I had a GitHub connection to ChatGPT that had been successfully configured and actively used for several weeks with a private repository.
At some point, ChatGPT could no longer access the repository.
The confusing part was that the GitHub plugin continued to appear installed/enabled. We had also previously attempted to re-enable/reconfigure it, which appeared successful, but actual repository access still did not work.
This created what looked like a configured connection without a functioning underlying authorization.
What we found
During troubleshooting on September 17, 2026, we went beyond checking whether the plugin was installed/enabled and tested an actual GitHub repository operation.
We eventually found that the underlying GitHub connection showed Connect rather than being operationally connected.
The recovery sequence was:
-
Open ChatGPT Settings → Plugins → GitHub.
-
Select Connect.
-
ChatGPT required MFA to be enabled before the connection process could continue.
-
Enable MFA on the OpenAI/ChatGPT account.
-
Complete the GitHub OAuth authorization.
-
Choose Only select repositories rather than All repositories.
-
Authorize the specific private repository.
-
Return to ChatGPT.
-
Test the connection with an actual repository read.
After this, ChatGPT successfully retrieved both private repository metadata and the contents of a file in the repository.
Important distinction
I do not know why the previously working authorization stopped working.
MFA was required during the successful reconnection process, but I do not have evidence that lack of MFA caused the original connection to fail.
The issue I can demonstrate is the difference between several states that can otherwise appear to mean “connected”:
Installed → Enabled → Provider Authenticated → Resource Authorized → Operational
In my case, Installed/Enabled did not mean Operational.
Because the integration had previously worked for weeks, this was particularly difficult to diagnose.
Suggested improvement
It would be useful if ChatGPT exposed connection health more explicitly—for example:
Installed | Authenticated | Authorized | Operational
A Test Connection button that performs a harmless provider operation would be even better.
For persistent projects and agentic workflows, it would also be valuable for ChatGPT to detect when a previously operational integration loses authentication or authorization and alert the user rather than leaving the failure latent until the integration is needed.
For anyone experiencing a similar problem: don’t rely solely on the plugin showing installed/enabled. Check the underlying connection and then verify it with an actual harmless operation against the provider.