MCP Tool Search Auto Mode — runtime test

MCP Tool Search Auto Mode is a Claude 2.1.7+ feature that defers MCP tool loading until runtime if tool descriptions would consume more than 10% of context, then discovers tools on-demand via the MCPSearch mechanism.

What the test attempted

The test design was sound: verify that when MCP tool metadata exceeds 10% of context, tools are NOT loaded upfront but instead discovered on-demand via MCPSearch. The payload would be significant—multiple MCP servers with many verbose tool descriptions—to trigger the threshold.

Result: INCONCLUSIVE

The test was not executed. The authors noted that achieving the 10% threshold requires "multiple MCP servers with many tools," a configuration most users don't deploy. Without a realistic heavy-MCP setup in place, the behavior could not be verified against the feature specification.

Why it matters

For heavy MCP users, this feature reduces context overhead: instead of all tools loading upfront (consuming context linearly with tool count), only discovery metadata stays resident. As MCP adoption scales, this becomes material. The 10% threshold acts as a reasonable circuit breaker—typical setups remain unaffected while power users benefit from graceful scaling.

Caveats

This remains untested at runtime. The 10% boundary itself is unverified. Actual MCPSearch latency and completeness on large tool inventories are unknown. Users with large MCP deployments would need to empirically confirm tool availability in the runtime, not just trust the feature exists.

Users can disable the feature by adding MCPSearch to disallowedTools in settings.

Status: Runtime test needed. Setup: heavy MCP servers with ≥10% context tool descriptions, then request an operation that would benefit from lazy tool loading.