MCP server processes never terminate, accumulating until OOM

[MariusWilsch]

Problem

MCP server processes spawned by metamcp never terminate after completing work, accumulating until memory exhaustion.

Quick Check

docker exec metamcp ps aux | wc -l
# If > 100, you have process accumulation

Evidence

Server 1 (14 days uptime):

• Process count: 50 → 2,018
• Memory usage: 213MB → 38.85GB
• ~140 processes accumulating per day

Server 2:

• 4.94GB/7.6GB RAM consumed
• OOM killer triggered repeatedly

Impact

• Each MCP operation leaves ~100MB zombie process
• Linear accumulation guarantees eventual OOM
• No cleanup mechanism exists

Expected vs Actual

• Expected: Process spawns → executes → terminates
• Actual: Process spawns → executes → sleeps forever

Environment

Ubuntu 24.04.2 + Docker + metamcp:latest

[zhjch05]

In this case I recommend you to try https://github.com/metatool-ai/metamcp/tree/docker-per-mcp

See related #128

I would also appreciate to see if it is the client that never terminate the streamable http connection correctly, because there is session clean up mechanism.

[zhjch05]

In the meantime I think metamcp should proactively terminate long sessions, because currently it relies on the delete endpoint of streamable http transport of MCP to free the session.

[MariusWilsch]

Could you elaborate a little bit of what you mean with the client that never terminates. Happy to share more data if I know what I'm looking for.

[zhjch05]

I wonder the MCP client side call the DELETE rest API of /mcp ever? If not every time it will create a new session without freeing existing ones.

[MariusWilsch]

I mostly use Claude code as MCP client. Do you think they are at fault?

[MariusWilsch]

Would you say that if I use SSE instead of http this issue would be resolved?

[zhjch05]

Super weird then. SSE may be better I think, so it clean up as long as the connection is cut off. But I am not sure how claude code spawn MCP connections, do connections always grow or not.

Anyways I will set a global lifetime to configure and auto cleanup these sessions when expiring.

[zhjch05]

And what I can share is that auto cleanup may break existing MCP connections. So I highly recommend you to use https://github.com/metatool-ai/metamcp/tree/docker-per-mcp

Because of this memory issue is happening for quite some time and I didn't find a good way to fix it. I will probably set docker-per-mcp branch as default to maintain soon.

[zhjch05]

2.4.15 gives an option to auto clean up session that passes lifetime, note that existing connection will see "Session not found" error and MCP client has to gracefully handle it and create a new session connection.