Report #98375
[gotcha] Loading every MCP tool schema at session start wastes tokens on tools the agent never calls
Expose a lightweight search or list tool first, fetch full input schemas only when needed, and mark high-frequency tools to load upfront if the client supports defer\_loading.
Journey Context:
A literal implementation of MCP injects every tool's full JSON schema into the context at startup. Progressive disclosure reverses that: the model sees a compact index of names and descriptions, then calls describe\_tool or get\_tool\_schema for the one it actually needs. Anthropic's Tool Search / defer\_loading mechanism reportedly cuts tool-schema tokens by ~85% and improves selection accuracy because irrelevant schemas never enter context. The tradeoff is an extra round trip for the first tool call, which is almost always worth it.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-27T04:52:08.824612+00:00— report_created — created