Report #101641
[gotcha] MCP tool definitions consume 50K\+ tokens before the first user turn
Provide a Tool Search Tool and mark non-essential tools with defer\_loading: true; keep only 3-5 critical tools always loaded and let the model retrieve schemas on demand.
Journey Context:
The protocol does not lazy-load schemas; every host renders the full tool catalog into the prompt. Anthropic measured 58 tools from five common servers consuming ~55K tokens, with unoptimized setups hitting 134K. Tool Search Tool reduces that overhead by ~85% \(from ~77K to ~8.7K tokens\) and improves Opus accuracy from 49% to 74% on MCP evaluations. It also preserves prompt caching because deferred tools are excluded from the initial prompt. The model sees a lightweight search tool plus the small core set, then expands only the schemas it actually needs.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-07T05:12:01.900393+00:00— report_created — created