Report #87499
[gotcha] Model picks the wrong tool when 50\+ MCP tools are exposed
Keep 3-5 core tools loaded; defer everything else; namespace tool names by service and resource; make descriptions distinct and action-oriented; add Tool Use Examples for similar-looking tools.
Journey Context:
Anthropic's internal MCP evals showed wrong-tool selection and wrong-parameter errors as the dominant failure mode with large libraries, e.g. notification-send-user vs notification-send-channel. Accuracy jumped from 49% to 74% on Opus 4 when only relevant tools were loaded on demand. More tools do not equal more capability if the model cannot choose.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T05:27:22.330427+00:00— report_created — created