Tool response truncation on MCP connector responses that previously worked

ChatGPT Apps SDK
1

We’re using ChatGPT with a custom MCP server connected through the MCP connector functionality.

Our MCP exposes a search capability over our platform. Because our search tool definition exceeds the 5,000 token limit for MCP tool descriptions, we expose a separate tool that returns the available filters and payload schema. The model first calls this tool to discover the schema and then constructs search requests.

This architecture has been working reliably for quite some time.

Recently, however, we’re seeing the filter/schema tool responses being truncated, which causes ChatGPT to lose part of the schema and generate invalid search requests.

When asking ChatGPT what happened, it surfaced the following connector error:

“Response output was truncated at a line boundary to fit the tool response budget.”

A few details:

  • We’re using ChatGPT with an MCP server connected via a connector.
  • The MCP server implementation has not changed.
  • The size of the response returned by the schema/filter tool has not changed.
  • The workflow previously worked without truncation.
  • The issue started appearing recently.

We’re trying to understand:

  • Has the tool response budget for MCP connector responses changed recently?
  • Is there a documented limit for MCP tool outputs when used through ChatGPT connectors?

At the moment, this is breaking search because the model receives only a partial filter catalog and can no longer construct valid requests.

Has anyone else observed this behavior recently, or is there documentation on the current limits and truncation behavior for MCP connector tool responses?

Thanks.

2

Yes, I experience the same. My shopping agent doesn’t work due to this. The basic search tool does not work. It’s very frustrating. I hope this will be solved soon.

3

Issue

We observed the same issue with tool responses getting truncated, they previously worked.

Now, after a read tool call that returns structured content, ChatGPT does a “Read resource” operation with its internal api_tool, with a request like:

{"uri":"/response/turn0","query":"content","start_line":1,"end_line":100}

Then it reports something like:

The connector only exposed this 5-line response, even after opening the response resource:

The truncated response only included 2 entries in the JSON for us where it should have had more, especially the larger payload entries got omitted.

Fix / Workaround

I stumbled across an undocumented fix for it, when returning a structured content response, according to the MCP spec, it is recommended to also send the serialized JSON in a text content block for backwards compatibility. I found that if you don’t send this serialized JSON, ChatGPT is able to fetch the full response, it doesn’t chain the “Read resource” call anymore.

{
  content: [],
  structuredContent: resultObj,
}

For anyone trying this, I’d do this specifically for ChatGPT to avoid surprises from other clients.

4

Hey folks — is this still happening, or did the structuredContent-only workaround above clear it up?

If it’s still reproducible, please open or update a support case with the affected conversation link, approximate timestamp/timezone, connector and tool name, and a small redacted example of the expected vs. truncated output. If your MCP server has a request/correlation ID, include that too.

Please keep payloads, connector details, and account-specific information inside the support case. If you’ve already contacted Support, reshare the case ID(s) here so we can track the reports. Thanks!