MCP Tool Search not auto-enabling when tools exceed 10% of context

Resolved 💬 5 comments Opened Jan 15, 2026 by amihos Closed Jan 18, 2026

Description

MCP Tool Search is announced as enabled by default when MCP tools exceed 10% of context, but it doesn't appear to be working.

References

Per Thariq Shihipar's announcement:

"Claude Code detects when your MCP tool descriptions would use more than 10% of context. When triggered, tools are loaded via search instead of preloaded."

Expected Behavior

When MCP tools exceed 10% of context, Tool Search should automatically enable and tools should be loaded on-demand via search rather than preloaded.

Actual Behavior

All MCP tools are still preloaded into context. Running /context shows:

MCP tools: 28.1k tokens (14.1%)

This is above the 10% threshold but tools are still fully loaded, not deferred.

Environment

  • Claude Code Version: 2.1.7
  • OS: Linux (WSL2)
  • MCP Servers: Multiple (Serena, Sentry, Context7, Playwright, Auggie, OpenMemory, IDE)

Steps to Reproduce

  1. Configure multiple MCP servers totaling >10% of context
  2. Start Claude Code
  3. Run /context command
  4. Observe MCP tools are fully loaded (28k+ tokens) instead of deferred

Questions

  1. Is there a setting required to enable this feature?
  2. Is ENABLE_TOOL_SEARCH=true an environment variable that needs to be set?
  3. Which Claude Code version includes this feature?

Related Issues

  • #12836 (Support Tool Search and Programmatic Tool Use betas)

View original on GitHub ↗

This issue has 5 comments on GitHub. Read the full discussion on GitHub ↗