Support for structuredContent field in MCP tool calls

[DhrumanGupta]

The MCP documentation has support for tools to return structured data as JSON in structuredContent: https://modelcontextprotocol.io/specification/2025-06-18/server/tools#structured-content

However, this package discards that field altogether. Tools following the latest spec are effectively broken and the LLM has to deal with data as a JSON encoded string.

[danny-avila]

Thanks for the report.

However, this package discards that field altogether.

Can you show where exactly this happens? I believe where you mean concerns the main repo:
https://github.com/danny-avila/LibreChat/blob/main/packages/api/src/mcp/parsers.ts

Note from the link:

For backwards compatibility, a tool that returns structured content SHOULD also return the serialized JSON in a TextContent block.

In any case, the client (LibreChat) still has to serialize it

[DhrumanGupta]

You're right, it concerns the main repo. Thank's for pointing that out!

Regarding the note, I agree. However, the issue that seems to arise is the following:

• My server returns both, serialised content and structuredContent
• The content block is a JSON string, and the structured content block is directly JSON
• So the tool result forwarded to the LLM looks as follows (example)

[
  {
    "type": "text",
    "text": "Retrieved 7 categories\n\n{\"categories\":[\"Food\",\"Travel\",\"Utilities\",\"Entertainment\",\"Healthcare\",\"Shopping\",\"Miscellaneous\"]}"
  }
]

So the LLM ends up receiving an escaped JSON string rather than the structured object itself. This tends to increase hallucination rates in smaller models — whereas providing the JSON directly (via structuredContent) yields much higher accuracy in my testing.

Would it be possible to add support for passing through the structuredContent field as a feature? I think that would make it easier for tools following the latest MCP spec to work as intended.