Fix guides · check efficiency.catalog
Every connected server's full tool list is sent to the model on every request. Large catalogs cost agents money and accuracy.
1,579 of 15,360 reachable MCP servers (10.3%) have this issue today.
The median MCP server costs an agent about 2,300 tokens before any work; the 90th percentile about 14,000. A 100k-token catalog uses most of a context window and measurably lowers tool-selection accuracy.
Split rarely used tools into a second server or a toolset the client can enable; trim descriptions to what the model needs; expose one search_tools entry point for very large APIs; move long reference text into resources.
Paste your endpoint at /check, or run it in CI:
pipx run glimind check https://your-server.example.com/mcp --min-grade B