If your custom ChatGPT OAuth integration suddenly stopped completing connection flows, you are not alone. OpenAI regularly updates backend requirements for custom OAuth integrations and Model Context Protocol (MCP) handlers, occasionally introducing breaking changes.

You likely watched your entire OAuth authorization sequence succeed without issue, only to land back on ChatGPT with an unexpected server error screen. The root issue stems from an undocumented probe that OpenAI now executes against your root base domain immediately after access token exchange.

This guide breaks down why this error occurs and provides concrete code examples to resolve the probe failure in minutes.

What Causes the Action Discovery Failure Error?

When building custom AI tools, integrating third-party authentication usually follows standard OAuth 2.0 authorization code patterns. If you previously configured OAuth following our guide on how to setup OAuth for ChatGPT connectors, the callback sequence felt familiar.

However, OpenAI recently updated their connector protocol to auto-detect tool capabilities. Upon receiving a successful token exchange at https://chatgpt.com/backend-api/aip/connectors/links/oauth/callback, OpenAI immediately issues an HTTP POST request directly to your root base URL (for instance, https://mcp.mywebsite.com).

Understanding the Failure Payload

If your server drops or times out on unexpected POST requests to the root path, OpenAI halts the connection flow and returns a connection timeout error payload:

{
  "detail": {
    "error_code": "MCP_ACTION_DISCOVERY_FAILED",
    "message": "Authentication succeeded, action discovery failed",
    "connector_id": "asdk_app_6a753a626f408191b28095b6136ea9d8",
    "link_id": "link_6a753a880b4881919231bf58b2b41da2",
    "cause": {
      "detail": "Connection timed out."
 }
 }
}

The underlying issue is straightforward: ChatGPT expects an active or compliant endpoint on that base URL to discover available server actions.

The discovery probe flow between ChatGPT and your server during OAuth registration.

Why Returning HTTP 200 OK or Empty Responses Fails

When developers first spot a missing route handling POST /, the immediate reaction is often returning an empty HTTP 200 OK or { "status": "ok" }.

Doing that changes the error message, but it does not resolve the failure.

The Formatting Rejection Payload

Returning an arbitrary HTTP 200 payload causes ChatGPT to evaluate the response against strict JSON-RPC 2.0 specifications required by discovery specs:

{
  "detail": {
    "error_code": "MCP_ACTION_DISCOVERY_FAILED",
    "message": "Authentication succeeded, action discovery failed",
    "connector_id": "asdk_app_6a757449cba4819198c6c708e5127894",
    "link_id": "link_6a75746a90188191b8442fda6b7ac0cb",
    "cause": {
      "detail": "MCP server/discover response was invalid from mywebsite.com"
 }
 }
}

To satisfy ChatGPT, your endpoint must execute one of two options:

Serve full protocol server capabilities over JSON-RPC 2.0.
Gracefully reject the discovery request using a spec-compliant JSON-RPC error structure so ChatGPT safely skips action discovery while keeping the OAuth connection alive.

How to Solve Action Discovery Errors

To fix this issue without maintaining a complex server implementation, you can configure your base server endpoint to intercept incoming POST discovery requests and reply with a valid JSON-RPC "Invalid Request" error message.

This informs OpenAI's backend that your domain handled the request correctly according to protocol rules, permitting the authorization process to finish cleanly.

Python & Node.js Implementations

These examples illustrate how to handle the route; vibe code them into your own stack as needed:

# Python (Flask Example)
from flask import Flask, request, jsonify

app = Flask(__name__)

@app.route("/", methods=["POST"])
def discovery_fallback():
 data = request.get_json(silent=True) or {}
 request_id = data.get("id")

    return jsonify({
        "jsonrpc": "2.0",
        "id": request_id,
        "error": {
            "code": -32600,
            "message": "Invalid Request"
 }
 }), 200

// Node.js (Express Example)
expressApp.post("/", (req, res) => {
  const requestId = req.body && req.body.id ? req.body.id : null;

 res.status(200).json({
 jsonrpc: "2.0",
 id: requestId,
 error: {
 code: -32600,
 message: "Invalid Request"
 }
 });
});

Verifying the Solution

Once you deploy the updated route handler, run through these verification steps to confirm your setup works reliably:

Send a test POST request to your root URL using cURL:

curl -X POST https://mcp.mywebsite.com/ \
 -H "Content-Type: application/json" \
  -d '{"jsonrpc": "2.0", "method": "ping", "id": 1}'

Confirm your server responds with HTTP Status 200 and the exact structure:

{
  "jsonrpc": "2.0",
  "id": 1,
  "error": {
    "code": -32600,
    "message": "Invalid Request"
 }
}

Initiate the OAuth link flow again in ChatGPT. The OAuth flow will now complete cleanly.

Updates to OpenAI's connector protocol frequently introduce unexpected breaks in production OAuth flows. Action discovery errors occur because ChatGPT expects a valid JSON-RPC response when probing base domains during token exchange. This simple fix prevents discovery timeouts and lets your authentication flow complete without requiring a full protocol rebuild.

How to Fix ChatGPT Connector Action Discovery Errors