There are two oAuth flows for MCP Authorization according to official modelcontextprotocol.
Which one is used by ChatGPT ?
- OAuth Flow of 2025-06-18
- OAuth Flow of 2025-03-26
Yes I’m a Plus user and I’ve created several mcp connections, mostly with n8n then API requests. Works fine. Just ask ChatGPT to send a JSON body with the request.
can’t see the option in my ChatGPT Team plan
Would love to see integration with Docker MCP toolkit. As far as i can tell this is not supported yet.
is there a tutorial to teach me how to set up oauth for a mcp server? I fail to figure it out by talking with chatgpt.
I tried to implement the oauth discovery like this
app.get("/.well-known/oauth-authorization-server/:resource", (_req, res) => res.json(AS_METADATA));
const AS_METADATA = {
issuer: ISSUER,
authorization_endpoint: `${ISSUER}/authorize`,
token_endpoint: `${ISSUER}/token`,
response_types_supported: ["code"],
grant_types_supported: ["authorization_code"],
code_challenge_methods_supported: ["S256"],
token_endpoint_auth_methods_supported: ["none"],
scopes_supported: ["openid", "profile", "email", "mcp"],
};
But chatgpt fails later, and I don’t know the meaning.
Failed to resolve OAuth client for MCP connector
can’t find the option in my ChatGPT Team plan as an admin role, the creation of `my connector` still returns {
“implements_retrievable”: false,
“reason_if_not_retrievable”: “search action not found”
}
I configured a streamable HTTP MCP endpoint with OAuth2.
Also getting
{“detail”: “Failed to resolve OAuth client for MCP connector: https://…/mcp”}
(in the Chrome console)
after the client was reading the .well-known/oauth-authorization-server endpoint
The server requires a ClientID + Secret (no dynamic registration), maybe this causes trouble (although it is defined in the MCP protocol that clients must support this).
I think the mcp session handling is currently incorrect.
According to the MCP specification (Session Management, 2025-06-18 spec version):
“An MCP ‘session’ consists of logically related interactions between a client and a server.”
Currently in chatgpt, however, MCP sessions are scoped only to the duration of a single tool call - and the session is deleted immediately afterward. I.e. **every call** is treated as its own isolated session, with no persistence across related interactions in the same conversation.
That doesn’t align with the MCP spec, idividual tool calls are clearly not the full scope of “logically related interactions.” A session should persist for the broader logical scope (e.g. a chat thread).
Because the session ID is discarded per tool call, it prevents us from enforcing proper session-level policies on the server side in our Open-Edison project, blocking us from protecting users from data exfiltration.
Open-Edison is a secure MCP control panel designed to address the “lethal trifecta” problem (See Simon Willisons blog on the topic ) - the dangerous combination of tool use, user secrets, and untrusted data that can enable catastrophic exfiltration attacks.
For our protection model to work, we need session IDs to persist across the full scope of related interactions, not reset per tool call.
Steps to Reproduce:
- Connect a custom MCP server to ChatGPT (we have steps in our github but any client with session support works)
- Invoke any tool multiple times within the same chat thread.
- Observe that each invocation is treated as a separate MCP session, with the previous session deleted immediately at the end of the call.
Expected Behavior:
MCP sessions in ChatGPT should persist for the logical scope of the conversation (e.g. a chat thread), in line with the MCP spec’s definition of “logically related interactions” and the requirement to reuse the `Mcp-Session-Id` across related requests.
Actual Behavior:
MCP sessions are reset and deleted for each tool call, effectively making every call its own session.
I hope this can be resolved.
So, I’ve got my MCP connected, and the connector interface successfully lists all the tools, and acknowledges that the connector is available for knowledge, however no actual chat sessions are able to use it. When attempting to query the connector the following is returned:
Can anyone offer any suggestions?
Cheers
Rich
Pretty cool! We were able to get it running and enable real time flight search as a POC. Let’s hope they open up an app store soon.
I have the same error for a tool call, which works perfectly in other AI agemts, like Claude web app, and n8n workflows.
Yes.
Settings → Connectors → Advance settings → enable developer mode
Then go back to Connectors tab, you will see create button after Browse connectors
I’m having the same issue with the tool calling. I tried explicitly referencing the tool names but I keep getting an error
Hi, Can you see the “Developer Mode” option when clicking on the Plus button anymore? My Plus account can no longer see it anymore from today (still see it on last Friday)
i have the same! Also worked for me on Friday and now no longuer see it in the tools section, of course chats no longer have access to it so just repeat from cached answers.
some has an explanation about this?
Hi Same for me, I have Plus subscription and cannot see my remote mcp server listed anymore in the list of sources in dev mode
Uhm, me too. Tried with some other MCP servers like PayPal, but no luck.
Same issue here! The option seems to have disappeared when you click the plus sign, worked last friday.
omg, had the same exact issue, emailed their support but its some annoying useless ai agent on the other side, ai fatigue is real
Yeah, it’s pretty bleak when it comes to support. Should really rename themselves OpaqueAI 
The error itself is quite simple - one of your tools is saying it returns an array, and that array can be empty, but you’re not actually specifying an empty array - so the parser bombs.
Have a look through your manifest and look for the anyOf element, then fix your tool appropriately.