MCP SSE Automatic Session Closure

[mmalinov]

Hello,

I am using LibreChat since a couple of days to create a proof-of-concept use case of an MCP server that I have created by using the JavaScript SDK. In short, it interacts with 3rd party API performing various tasks.
Since the current version of the MCP Server is a mock-up version, there are some issues with the session handlings for the messages.
In short, when I start the server and LibreChat connects to it, I can use the chat and the tool calling for exactly 5 minutes, after which LibreChat automatically closes a session and initiates a new one. However, with the new session the tool execution timeouts on the client side(I can see from the server and 3rd party API that the requests are executed). When I restart LibreChat, everything again works for 5 minutes.

Here is my question: since this is only a proof-of-concept implementation, I don't really care about proper session handling/closure. How can I make the LibreChat threshold of automatic session closure longer or remove it at all?

I run everything in docker and currently only have the docker-compose.override.yml to include the librechat.yml.

Thanks!

[danny-avila]

Hi thanks for the report.

Can you clarify a few things:

• Is the tool call running a long time, then timing out at 5 minutes?
• Or, you are using the MCP server tools intermittently, but experiencing issues after 5 minutes?

In any case, the only "inactivity" for MCP user sessions are hardcoded to 15 minutes, but this would not affect you while using the tool.

[danny-avila]

However, with the new session the tool execution timeouts on the client side(I can see from the server and 3rd party API that the requests are executed)

What do you see exactly that is timing out? Is the web app "timing out"?

When I restart LibreChat, everything again works for 5 minutes.

You restart the server, or refresh the page?

[mmalinov]

Hi Danny,

The tool execution time is very short, a couple of seconds.
I can use the MCP server tools constantly for 5 minutes since the start of the LibreChat server. Afterwards I have to restart the server.
I believe that everything is related to the first MCP session when LibreChat is starting and connecting to the server.
Here are some logs from the MCP Server and screenshots from LibreChat:

MCP Server listening with HTTP/SSE transport on port 3001
SSE Endpoint: http://localhost:3001/mcp/message
Message Endpoint: http://localhost:3001/mcp/message?sessionId=<sessionId>
GET /mcp/message - Incoming request from ::ffff:172.22.0.6
Request Headers: {
  "host": "10.20.20.88:3001",
  "connection": "keep-alive",
  "accept": "text/event-stream",
  "accept-language": "*",
  "sec-fetch-mode": "cors",
  "user-agent": "node",
  "pragma": "no-cache",
  "cache-control": "no-cache",
  "accept-encoding": "gzip, deflate"
}
SSE connection established from ::ffff:172.22.0.6 for session: **3ed3fea6-7087-452f-aeff-bc80372414a1** 

The connection with sessionId 3ed3fea6-7087-452f-aeff-bc80372414a1 is established when I start LibreChat. I can see all tools from the MCP.
Afterwards, I start sending messages and interact with the tools without any problems, but a new session is established:

SSE connection established from ::ffff:172.22.0.6 for session: 3f1f2bc9-aea0-4763-ac22-d5946c73788e
Response Headers (SSE Handshake): {
  "x-powered-by": "Express"
}
MCP Server connected for session: 3f1f2bc9-aea0-4763-ac22-d5946c73788e
Received POST message for session: 3f1f2bc9-aea0-4763-ac22-d5946c73788e
[MCP Tool: Get current user profile] Attempting to create URL from: http://testenv.example.com/api/v1/client-users/profile
POST /mcp/message - Incoming request for session 3f1f2bc9-aea0-4763-ac22-d5946c73788e from ::ffff:172.22.0.6. Query: {"sessionId":"3f1f2bc9-aea0-4763-ac22-d5946c73788e"}

At exactly 5 minutes, the initial session is closed and LibreChat stops seeing the responses from the MCP Server:
SSE connection closed for session: 3ed3fea6-7087-452f-aeff-bc80372414a1

Received POST message for session: 3f1f2bc9-aea0-4763-ac22-d5946c73788e
[MCP Tool: Get project options] Attempting to create URL from: http://testenv.example.com/api/v1/projects/options
SSE connection closed for session: 3ed3fea6-7087-452f-aeff-bc80372414a1
GET /mcp/message - Incoming request from ::ffff:172.22.0.6
Request Headers: {
  "host": "10.20.20.88:3001",
  "connection": "keep-alive",
  "accept": "text/event-stream",
  "accept-language": "*",
  "sec-fetch-mode": "cors",
  "user-agent": "node",
  "pragma": "no-cache",
  "cache-control": "no-cache",
  "accept-encoding": "gzip, deflate"
}

A new one is automatically created, but the LibreChat agent continues to use the second session(which is still not closed):
MCP Server connected for session: 92ff3f6c-c576-44a0-9703-50928ec2e78a
Received POST message for session: 3f1f2bc9-aea0-4763-ac22-d5946c73788e
[MCP Tool: Create a new project] Attempting to create URL from: http://testenv.example.com/api/v1/projects

This is how things look like in LibreChat FE:

The tool execution "Error" is exactly when the first session is closed 3ed3fea6-7087-452f-aeff-bc80372414a1:

However the tool execution is successful on the MCP Server side:
Received POST message for session: 3f1f2bc9-aea0-4763-ac22-d5946c73788e
[MCP Tool: Create a new project] Attempting to create URL from: http://testenv.example.com/api/v1/projects

After I restart the whole LibreChat docker stack, everything works smoothly again for the first 5 minutes(e.g. as long as the initial MCP session is alive)

Thanks for the support!

[danny-avila]

Are you on the latest version of LibreChat? At one point, I removed automatic pings from LC to the MCP server, but then I added it back in due to connection issues, since it helped keep the session alive

Updating Instructions (docker):
https://www.librechat.ai/docs/local/docker#update-librechat

It seems you're not alone in this (independent of LibreChat):
home-assistant/core#140667

Any LC logs when you notice the disconnect would be helpful, too.

[danny-avila]

I believe this issue is specific to the Eventsource SSE protocol since I find more hits when searching 5 minute timeouts for that than for anything MCP-related.

Are you receiving any pings from LC? Are you sending anything back for pings to keep the connection active?

You may also want to look into refactoring the server with the streamable http transport, which LC will support soon, which is meant to address caveats of long-lived connections like this, and is more suitable for a multi-user environment:

#7020
https://modelcontextprotocol.io/specification/draft/basic/transports#streamable-http

[danny-avila]

Example SSE-related timeouts at exactly 5 minutes:

• https://stackoverflow.com/questions/71207873/sse-connection-keeps-failing-every-5-minutes
• How do you keep the SSE connection open longer than 5 mins? pocketbase/pocketbase#3054
• https://forum.cursor.com/t/sse-timeout-error-after-5-minutes/82747

I found nothing in the MCP TS SDK related to a 5 minute timeout

[mmalinov]

Hi Dani,

Thank you very much for the responses! It was indeed related to SSE itself, rather than LC or the MCP TS SDK.
The LibreChat instance is at latest commit on 04-05-2025 on the main branch, I haven't updated since.

I implemented automatic ping directly in the MCP Server and now the sessions are not expiring and everything works as expected.
I will definitely refactor the MCP to support StreamableHTTPTransport in the future.

Thanks again!

[danny-avila]

Awesome, great to hear!