Realistic number of tools loaded at the same time?

[frippe75]

Hi everyone,

I'm working with a large OpenAPI 3.0 spec (~300 paths) and using FastMCP.from_openapi to dynamically generate tools based on groups of paths. This is part of a hierarchical discovery pattern where the LLM can request to "activate" a capability group.
I'm running into a significant performance issue. When a tool is called that triggers the creation of a new group of tools from the spec, the server process maxes out its CPU and becomes unresponsive for 30+ seconds. A simple performance test isolating the FastMCP.from_openapi call confirms that this function is the bottleneck.
My question is: Is this a known performance characteristic when generating a large number of tools (e.g., 50-100) dynamically during a request?
My current plan is to pre-generate all the tool groups at server startup and cache them, which I suspect is the right architectural pattern. I just wanted to confirm if this is the expected behavior or if there might be a more efficient way to handle dynamic tool loading from a large spec.
Thanks!

[sandeepkunusoth]

H were u able to solve this issue?