I’m encountering the same 424 “Failed Dependency” error in the OpenAI Agent Builder when connecting my custom MCP server.
My MCP server responds correctly (handshake/initialize returns 200, list_tools/list_actions returns valid JSON) when tested via direct API calls. It is publicly reachable, over HTTPS, and the schema matches the documented MCP spec.
However, when Agent Builder invokes the internal endpoint (e.g., /v1/dashboard/responses/mcp/list_actions), it fails with 424 – suggesting the problem lies in Agent Builder’s internal validation or workflow orchestration, not in my server’s implementation.
Some helpful reproduction hints:
The error appears consistently at the tool-listing / list_actions stage.
The handshake (initialize) is successful but the next step fails silently (server may not even receive the request).
I hope the OpenAI team can investigate the internal lifecycle between initialize → list_actions and clarify whether there are undocumented constraints (e.g., naming conventions, response times, header requirements) that could trigger the 424. Any update or workaround would be greatly appreciated.
@Foxalabs Thanks for the update, but can I please quickly ask: am I missing something or is there a workaround that I’m not aware of?
It seems that this is quite a big blocker for anyone trying to implement AgentBuilder workflows with their own custom functionality via the MCP route? And so soon after the big launch, I’d assume this would be hotfixed a week ago (sorry for being blunt).
My thoughts exactly, unless I’m missing something MCP servers are the only way to have your agent interact with the outside world. So if that doesn’t work, what are the use cases left ?
I just wanted to add my case here since I’m experiencing the exact same issue with all MCP servers in Agent Builder (Zapier, Google Drive, and Google Sheets).
Everything was working fine until recently — then all my connectors suddenly started failing with the same message:
{"jsonrpc":"2.0","id":null,"error":{"code":-31996,"message":"None of the security schemes (mcpBearerAuth) successfully authenticated this request."}}
My setup uses multiple MCP endpoints, all reachable and working correctly when tested manually (the authentication token is valid, and the endpoint returns the proper tool list when called directly).
The error only happens inside Agent Builder, suggesting that the platform has stopped forwarding the MCP bearer token (mcpBearerAuth) to the remote server.
I’m a beginner still learning how Agent Builder and MCP integration work, so I triple-checked everything:
Regenerated tokens and reconnected all servers.
Tested multiple accounts and workspaces.
Tried creating brand new workflows from scratch.
Confirmed no permission or role restrictions were changed.
Everything points to a platform-side regression or authentication bug that broke MCP connections simultaneously across all connectors.
Thank you to everyone contributing here — I really appreciate the community helping new builders like me figure this out.
Hi everyone! I’m experiencing the same problem with both new and existing MCP connections. The one that was working fine yesterday now shows ‘Unable to load tools,’ and the same issue occurs with new connections.
Good to know that I am in not alone in this. If OpenAI wants to lock-in serious developers, they have to prioritize stability over feature expansion. Their maturity model is way below best practices.
Authorization → Bearer sk-proj-p i have used this approach as well, but LLM doesn’t call this tool when you asked in the chat. It shows the tools but doesn’t work.
My MCP server responds correctly (handshake/initialize returns 200, list_tools/list_actions returns valid JSON) when tested via direct API calls.
However, when Agent Builder invokes the internal endpoint (e.g., 
