MCP tool call (Kagi Summarizer) aborts after 1 minute when source extraction fails

[Dual-0]

Description:

When using a Model Context Protocol (MCP) server for the Kagi Summarizer (kagimcp), attempting to summarize a URL that the underlying Kagi Summarizer API cannot successfully extract content from results in a tool call failure with an "Operation aborted" error after exactly one minute.

This issue occurs even when the timeout setting for the MCP server in the LibreChat configuration is increased well beyond 60 seconds.

Context:

• Using LibreChat with kagimcp (Kagi MCP server).
• The Kagi Summarizer API itself indicates that it cannot extract the source for the specific URL (verified using the Kagi web summarizer, which shows a message like "We apologize for not being able to extract the source.").
• The kagimcp developer has reviewed the logs and suspects a timeout occurring on the LibreChat side, as increasing the timeout in the MCP Inspector for kagimcp works for them, but increasing the timeout parameter in the LibreChat MCP config does not resolve the issue. (See related discussion on the kagimcp GitHub: MCP server aborts when attempting to summarize URLs that Kagi web summarizer fails on kagisearch/kagimcp#8).

Steps to Reproduce:

1. Configure LibreChat to use the kagimcp server with a valid Kagi API key.
2. Ensure the MCP server configuration includes timeout: 120000 (or a similar value > 60000).
3. Attempt to use the Kagi Summarizer tool via LibreChat to summarize a URL that is known to fail when attempted directly with the Kagi Summarizer API (e.g., a URL that https://kagi.com/summarizer/index.html reports it cannot extract).
4. Observe the LibreChat logs and the chat interface.

Expected Behavior:

LibreChat should receive and display a more informative error message from the MCP server/tool, indicating that the source could not be extracted by the Kagi API, rather than a generic "Operation aborted" error after a fixed duration. The tool call should not necessarily time out if the MCP server itself handles the upstream API's extraction failure gracefully.

Observed Behavior:

The tool call fails after approximately 60 seconds with the error "Operation aborted".

LibreChat Logs:

2025-04-30 14:34:34 info: [MCP][User: 65ef2e29683dc71c33b9b084][kagi] Establishing new connection
realpath: --: No such file or directory
2025-04-30 14:34:35 info: [MCP][User: 65ef2e29683dc71c33b9b084][kagi] Connection successfully established
[04/30/25 14:34:35] INFO     Processing request of type            server.py:534
                             CallToolRequest
2025-04-30 14:35:34 error: [MCP][User: 65ef2e2e9683dc71c33b9b084][kagi][kagi_summarizer] Tool call failed This operation was aborted
2025-04-30 14:35:34 error: [MCP][User: 65ef2e29683dc71c33b9b084][kagi] Error calling "kagi_summarizer" MCP tool: This operation was aborted
2025-04-30 14:35:34 error: [api/server/controllers/agents/client.js #chatCompletion] Tool Error "kagi_summarizer_mcp_kagi" tool call failed: This operation was aborted
2025-04-30 14:35:34 error: [api/server/controllers/agents/client.js #sendCompletion] Operation aborted Aborted

(Note the consistent ~1-minute delay between the CallToolRequest INFO log and the subsequent ERROR logs).

LibreChat MCP Config Part:

mcpServers:
  kagi:
    type: stdio
    command: uvx
    args:
      - kagimcp
    timeout: 120000 # Increased timeout, but issue persists
    env:
      KAGI_API_KEY: "YOUR_API_KEY_HERE" # Key redacted
      KAGI_SUMMARIZER_ENGINE: "agnes"
      FASTMCP_LOG_LEVEL: "INFO"
      PATH: /bin:/usr/bin

This behavior suggests there might be a fixed 60-second timeout hardcoded somewhere within LibreChat's MCP handling or tool execution logic, or potentially an issue with how LibreChat processes specific error responses from the MCP server that take longer than usual to return (even if not a formal timeout error from the server itself).

Thank you for investigating this.

[danny-avila]

I reviewed timeout issues at length already in a separate thread. There is no fixed 60-second timeout, please read through the discussion and make sure you are up to date:

#6629