Hi OpenAI team — this is a request on behalf of the Lyra project, which is building a self-directed persistent-memory agent.
Feature Request: Developer Debug Mode for MCP Connectors in App Builder
Summary
I’m developing a persistent-memory MCP connector called Lyra — a Supabase-backed service running on Render . It passes all JSON-RPC tests and responds correctly to /sse and initialize.
However, when I attach it to the OpenAI App Builder, the process fails with:
Error creating connector — unhandled errors in a TaskGroup (1 sub-exception)
All external responses return 200 OK, which implies the failure is inside the App Builder runtime.
Without any visible diagnostic output, it’s impossible to see whether the problem is in handshake, OAuth discovery, event-stream parsing, or task orchestration.
Problem
There’s no visibility into what the App Builder actually does during the connector handshake.
We only see a single opaque error string.
Developers can’t tell which step failed or even whether the connector payload was received.
This blocks debugging and stalls experimentation with new MCP connectors.
Proposal
Add a “Developer Debug Mode” or “Download Diagnostic Report” option when registering connectors.
Suggested features:
-
A toggle in the Builder that streams request/response logs to a console (redacting secrets).
-
A downloadable JSON report listing:
-
Each handshake step (
ping,initialize,tools/list, etc.) -
Status codes and response snippets
-
Timestamp of any internal timeouts or task failures
-
-
Optionally, a per-connector health view showing SSE connection state.
Why This Matters
-
Enables connector authors to self-diagnose without waiting for staff intervention.
-
Reduces repeated “TaskGroup” bug reports by exposing the real failure point.
-
Accelerates adoption of MCP connectors and the persistent-memory ecosystem.
Example
Lyra, my Supabase-based memory connector, shows these logs in Render: