I Measured 27,257 MCP Connections. The p90 Session Waits 35s.

In July, a developer created a shell script called mcp-optional to remove two standard input/output (stdio) servers from their configuration. These servers were causing memory issues by spawning their own Node process per session, leading to increased RAM usage and swap thrash. The developer disabled these servers by default and re-enabled them on demand, which resolved the memory problem.

However, upon reviewing the MCP connection logs left by Claude Code on the developer's laptop, they discovered that one of the two servers was not only causing a significant RAM tax but also the slowest connection point by a wide margin. The developer measured the connection latency over 35 days, analyzing 33,599 log files and 27,257 successful connection pairs across 22 servers.

While the median connection time appeared acceptable at 582 milliseconds, 39% of connections took longer than a second, with a worst-case scenario (p99) of 14 seconds. The issue was traced back to OAuth token refreshes, which added an additional 2.6 seconds to connection times despite the server's response time being only 550 milliseconds.

This was due to the client having to obtain a new access token during connection, which significantly impacted the overall connection speed. Another factor contributing to the slow connections was the use of stdio instead of HTTP transport, which was 3 times faster in the median but worse in the tail. The developer also noticed that multiple connections to different servers led to longer connection times, with a median of 8 servers connected simultaneously and a p90 of 35.5 seconds.

A significant 29.7% of connection attempts failed to establish a connection at all, with one integration being responsible for 7,422 of these failed connections. The developer capped their measurements at the client's declared 30-second timeout, removing a small percentage of slow connection samples that were unrelated to actual server issues.

Ultimately, the developer concluded that counting the number of servers before tuning them is crucial, with a median of 8 servers and a p90 of 35 seconds. Removing a mediocre server can significantly improve performance compared to optimizing a good one, and opting for heavy servers to be enabled only when necessary can lead to the highest leverage improvement.