Sporadic error when calling an MCP tool - LangChain related

[kapustor]

What happened?

Sometimes when calling an MCP tool with Sonnet 3.7 (doesn't depend on exact MCP tool) there is an error:

Error processing tool: Tool response format is "content_and_artifact" but the output was not a two-tuple.\nResult: "list_tables MCP server tool call failed."

This error originates from Langchain's check located here.
Though originally this is a problem on the Langchain side, it seems it can be fixed on the LibreChat side by creating a wrapper here:

      const formatted = formatToolContent(result, provider);
      if (!Array.isArray(formatted)) {
        this.logger.warn(`${logPrefix}[${toolName}] Tool response not an array (type: ${typeof formatted}). Wrapping as [content, undefined].`);
        return [formatted, undefined]; // Use undefined instead of null to match type
      } else if (formatted.length !== 2) {
        this.logger.warn(`${logPrefix}[${toolName}] Tool response array has ${formatted.length} elements instead of 2. Fixing format.`);
        if (formatted.length === 0) {
          return [`${toolName} returned no results`, undefined];
        } else if (formatted.length === 1) {
          return [formatted[0], undefined]; // Add undefined as second element
        } else {
          return [formatted[0], formatted[1]]; // Just keep first two elements
        }
      }
      return formatted;

Version Information

d8337e0

Steps to Reproduce

Unfortunately, I didn't find any reliable way to reproduce this error, but it appears quite often.

What browsers are you seeing the problem on?

No response

Relevant log output

Screenshots

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

The format used is intentional and allows for multi-modal outputs from tools. certain providers/models will cause this error, because they cannot accept them.

When you say anthropic's Sonnet 3.7, do you mean the LLM behind a custom endpoint or direct? Some exact details on your LLM setup would help troubleshoot what's going on.

The suggestion you made should not be made because formatToolContent is already taking care of this from a provider standpoint.

[danny-avila]

Ok, I can see that this is happening any time the MCP tool throws an error, for which the response format is not being returned correctly. I will handle this with better logging.

@kapustor this is likely happening intermittently for you because of some other error, for which you can find in the logs for exact details. You can find the exact details on finding the logs here: https://github.com/danny-avila/LibreChat/issues/new?template=BUG-REPORT.yml

@TiMoMuc you're likely encountering this "every time" due to some config error with postgres. I will try to get my config working on my end so I can suggest a solution.

[danny-avila]

@TiMoMuc now postgress is working for me with a proper config:

Using common defaults for postgres:

mcpServers:
  postgres:
    command: "npx"
    args:
      - "-y"
      - "@modelcontextprotocol/server-postgres"
      - "postgresql://myuser:mypassword@localhost:5432/mydatabase"

[kapustor]

Thank you so much @danny-avila ! This makes sense.

I will handle this with better logging.

Does this mean you will be able to fix it on the LibreChat side, so if an MCP tool throws an error for which the response format is not correct LibreChat will fix it?

[danny-avila]

@kapustor Currently, the response format for errors is being returned incorrectly. I will fix that. There will still be an "error" from the MCP response.

The MCP Server could still be an incorrect format, or an unhandled format, and it would be a separate issue, but this change will help narrow it down if it's truly something else.

You didn't provide the logs for the exact error you're having, or we could determine that now.

Error processing tool: Tool response format is "content_and_artifact" but the output was not a two-tuple.\nResult: "list_tables MCP server tool call failed."

This is what you see in the tool output and is not a log from the server

[TiMoMuc]

@danny-avila Thank you as well! The hint, that the problem could be the connection to the database was very helpful. I had to point to my host network through using host.docker.internal.

  postgres:
    command: "npx"
    args:
      - "-y"
      - "@modelcontextprotocol/server-postgres"
      - "postgresql://myuser:[email protected]:5432/mydatabase"

Now it's working :)

[danny-avila]

Closed by #6687