Report #3552
[gotcha] MCP tool schemas consume 10-50K tokens before the first user message
Instrument tokens per tool manifest, keep each tool description under ~500 tokens, and defer rarely-used tools via defer\_loading or a search→describe→execute gateway instead of loading full schemas every turn.
Journey Context:
Auto-generating tools from whole OpenAPI specs is the classic trap: every name, description, and JSON Schema property is repeated into the model's context on every turn. A 106-tool MySQL MCP server was measured at 207 KB / ~54,600 tokens on every initialization even when only 2-3 tools were needed. Verbose descriptions can help accuracy, but the tax grows linearly and quickly crowds out project context. The right call is to measure first, curate aggressively, and use progressive disclosure. Anthropic's Tool Search beta reports ~85% token reduction by loading only matched tools.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T17:32:17.649519+00:00— report_created — created