The first call is the slow one
Average cold start penalty across the catalogue is 31.6 ms. The worst we recorded is 14,586 ms. If your MCP server sleeps, your agent waits.
We measure two things on every round: how long a warm server takes to answer, and how much extra the very first call costs. The gap between those two numbers is where serverless hosting shows up.
The numbers
average modern latency 770.2 ms
average cold start extra 31.6 ms
worst cold start extra 14,586 msThe average is comfortable. The tail is not. Nearly fifteen seconds of extra wait on a first call is long enough that most agent frameworks give up before the answer arrives.
Why it happens
Plenty of MCP servers are deployed as scale-to-zero functions. That is a sensible default for a side project — you pay nothing while idle. But an agent does not retry patiently. It sends one call, waits for its timeout, and treats the silence as a broken tool.
If you host one
- Know whether your platform scales to zero — many do it by default.
- Keep one instance warm if anything automated depends on the server.
- Measure the first call after an idle period, not the tenth call in a loop.