This error appeared to be a red herring.
Yesterday the UI surfaced a “Not Safe” error when trying to create my connector.
There are obviously checks for naming conventions or something, and one of my tools could have been inferred to allow execution of code (it didn’t).
I renamed the tool to something more benign, and it happily connected it.
So RunPythonScript will fail, RunBridgeTask doesn’t.
A few other “Gotchas”:
Make sure you are only exposing single tool names - sounds dumb but copy-pasting code I didn’t change the method name and description, so the bridge saw 2 of the “same” tools.
Make your method names and descriptions… Descriptive. PutFile(“Writes a file”) seems to trigger alarms, FilesPutInAllowedDirectory(“Writes a file to an approved directory”) is ok.
Don’t have similarly named tools - “PutFile”, “PutFiles” seems to make it antsy.
Keep naming scoped - “FilesPutInAllowedDirectory”, “FilesDeleteFromAllowedDirectory” appear as files_put… files_delete…
Don’t trust that because the MCPInspector and OAI tools list ok mean the tool is usable. In the above examples, it happily picked the tools up, but refused to use some of them, even though they were valid and functional.
Avoid using terms like “Raw” or “Unknown” in your tool names.
My custom MCP connector (OAuth + Flask) was fully functional up to Oct 7 2025.
It still returns 200 for initialize and tools/list (11 tools total) when queried directly,
but ChatGPT Developer Mode now shows “something went wrong” and won’t list or call tools.
Confirmed from PowerShell logs; seems identical to others’ reports. Dose anyone know if OpenAI is working or knows of this problem?
FOUND IT OK OK WTF is super complicated and i s very very hard to find XD and if you dont creatre befor a connector from chat you cant access to developer mode in the workspace :S
For those looking for it. Once in developer mode, you need to re-enable developer mode (?) which will allow you to select the dev mode MCP connector. These connectors still are not available for non-workspace admins for team/ent accounts it seems unless someone wants to prove me wrong.
Although all of my tools are read-only ( mentioned that in tool description as well ) I always end up with “Error creating connector - Connector is not safe”. Using no authorization. Would that me mandatory?
Today, to use my MCP tool, chatgpt is showing this popup. After I clicked on the Continue button, it says that the tool is now connected, and when I ask again to use the tool, Chatgpt show the same popup again. OpenAI is pushing changes without testing it. This is so annoying!
I am trying to make this MCP work. It briefly worked yesterday, but looks like OpenAI is deploying new features every hour and they are not testing them. There is no communication. And no shame.
