In `litellm/proxy/_experimental/mcp_server/mcp_server_manager.py`, `get_prompts_from_server` / `get_resources_from_server` call `client.list_prompts()` / `client.list_resources()` on **every** aggregate `/mcp` request against **every** backing server — including servers that don't implement those optional MCP methods at all. Confirmed against 6 different real MCP servers, all of which raise "Method not found" on every single call: 2 wasted round-trips per unsupporting server, on every gateway request.

The file already has the right pattern for this elsewhere — `_upstream_initialize_instructions_probed_at`, a per-server monotonic-timestamp cooldown used for the initialize-instructions probe — it's just not applied to these two methods.

**Suggested fix:** extend the existing cooldown-cache pattern to `get_prompts_from_server` / `get_resources_from_server`, e.g. a per-server cooldown once a server has responded "unsupported."