MCP Servers Context window issue

[zl-Diamon]

What happened?

We found that when the amount of data returned by mcpServer is small, it works normally. However, when the data volume is large, it reports an error: Something went wrong. Here's the specific error message we encountered: An error occurred while processing the request: 400 status code (no body).

Version Information

Steps to Reproduce

If you directly use mcpServer to return a large amount of data, this error will occur. However, when mcpServer returns a small amount of data, it can be parsed and returned correctly.

What browsers are you seeing the problem on?

Chrome

Relevant log output

error: [api/server/controllers/agents/client.js #sendCompletion] Operation aborted 400 status code (no body)
error: [api/server/controllers/agents/client.js #sendCompletion] Unhandled error type 400 status code (no body)

Screenshots

No response

Code of Conduct

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

[danny-avila]

It's likely a context window issue, what provider and model are you using?

You may get around it if you set an appropriate value for max context tokens:

Potentially related: #7818 (comment)

[zl-Diamon]

I was using the Qwen3-14B-AWQ model. I set the maximum value given to me by deepwiki, but it still failed

[zl-Diamon]

How should I deal with it

[danny-avila]

Set only the max context amount to 32k for testing (the model likely only supports 32k), and nothing for max output tokens (leave blank).

If you still get an error, you can share more logs for when it happens.

Log locations:

• Docker: Project root directory ./logs
• npm: ./api/logs

There are two types of logs that can help diagnose the issue:

• debug logs (debug-YYYY-MM-DD.log)
• error logs (error-YYYY-MM-DD.log)

Error logs contain exact stack traces and are especially helpful, but both can provide valuable information.

[zl-Diamon]

I set the max context amount to 32k,But It's always this mistake.
debug.log:

error.log:

[randomhack-com]

I also get this error using a custom OpenAI endpoint. I can chat with the endpoint directly but as soon as I use the endpoint in an Agent or enable search I get this error when attempting to run any completion.

agent / model config:

UI output:

container log:

2025-06-18 15:39:17 error: [api/server/controllers/agents/client.js #sendCompletion] Operation aborted 400 status code (no body)
2025-06-18 15:39:17 error: [api/server/controllers/agents/client.js #sendCompletion] Unhandled error type 400 status code (no body)

debug log:

2025-06-18T15:39:17.343Z debug: [BaseClient] userMessage
{
  messageId: "a6d350e4-a608-43e1-bb9e-76a8990323f2",
  parentMessageId: "00000000-0000-0000-0000-000000000000",
  conversationId: "f4d4bc28-16d0-4ed5-b625-0483c2e60616",
  sender: "User",
  text: "hello",
  isCreatedByUser: true,
  tokenCount: 5,
}
2025-06-18T15:39:17.409Z debug: [AgentController] Request closed
2025-06-18T15:39:17.414Z debug: [AgentController] Performing cleanup
2025-06-18T15:39:17.414Z debug: [AgentController] Cleaning up abort controller
2025-06-18T15:39:17.415Z debug: [AgentController] Cleanup completed

[xyxxxxx]

I encountered the same error. The endpoint is a vLLM deployment of the Qwen3-8B model. If the MCP server is not enabled, the model responds normally. But once the MCP server is enabled, it immediately returns a 400 status code (no body). The vLLM server logs show:

INFO: 10.233.67.207:0 - "POST /v1/chat/completions HTTP/1.1" 400 Bad Request.

Unlike the issue mentioned above, this has nothing to do with the size of the data returned by the MCP server — as long as the MCP server is enabled, even a simple input like "hello" will trigger the error.

LibreChat version: v0.7.9-rc1

[xyxxxxx]

p.s. the request payload in browser that trigger the error:

{"text":"hello","sender":"User","clientTimestamp":"2025-07-09T17:00:18","isCreatedByUser":true,"parentMessageId":"d9309524-01f7-4bd1-a53d-f21ef72bf368_","conversationId":"072aad3a-2d47-4f17-8984-46b35a457d09","messageId":"4b27fc15-9293-4249-aecd-1190c1c63c48","error":false,"responseMessageId":"d9309524-01f7-4bd1-a53d-f21ef72bf368__","endpoint":"vLLM","endpointType":"custom","model":"Qwen3-8B","modelLabel":"law-assistant","promptPrefix":"","temperature":0.2,"resendFiles":true,"greeting":"","spec":"law-assistant-qwen3","maxContextTokens":32768,"max_tokens":32768,"key":"never","ephemeralAgent":{"mcp":["law-searcher"]},"isContinued":false,"isTemporary":false}

the request payload that is ok:

{"text":"hello","sender":"User","clientTimestamp":"2025-07-09T17:00:30","isCreatedByUser":true,"parentMessageId":"d9309524-01f7-4bd1-a53d-f21ef72bf368__","conversationId":"072aad3a-2d47-4f17-8984-46b35a457d09","messageId":"cbbb5c77-8a1e-4850-b494-968861afd046","error":false,"responseMessageId":"d9309524-01f7-4bd1-a53d-f21ef72bf368___","endpoint":"vLLM","endpointType":"custom","model":"Qwen3-8B","modelLabel":"law-assistant","promptPrefix":"","temperature":0.2,"resendFiles":true,"greeting":"","spec":"law-assistant-qwen3","maxContextTokens":32768,"max_tokens":32768,"key":"never","ephemeralAgent":{"mcp":[]},"isContinued":false,"isTemporary":false}

It seems that the only difference locates in mcp.

I set up modelSpecs in librechat.yaml.