Lazy-Load MCP Definitions Cuts Token Overhead 85% — 51K to 8.5K Tokens Automatically

Claude Code's Tool Search feature — shipped January 2026, now on by default — solved one of the most expensive hidden costs in agentic development. Previously, connecting 7 MCP servers consumed approximately 51,000 tokens of context just loading tool definitions before any work began. Tool Search defers all definitions and injects a single search tool instead, dropping overhead to ~8,500 tokens — an 85% reduction with zero configuration. Joe Njenga / Medium

The mechanism: when Claude needs a tool, it searches by keyword and selectively loads 3–5 relevant definitions per query. The feature auto-activates when total tool descriptions exceed 10K tokens. For builders running MCP-heavy setups — database servers, browser automation, file management, custom business logic — this is a pure context budget recovery play. Those 42,500 recovered tokens are now available for actual work: more code in context, longer conversation history before compaction, deeper retrieval results.

Combined with path-scoped .claude/rules/ files that defer instruction loading until matching files enter context, and CLAUDE.md kept under 200 lines with lazy pointers to module-specific sub-files, the full token discipline stack recovers the majority of context budget that naive configurations waste. Claude Code Docs If you haven't audited your MCP token footprint, run /context in Claude Code — the new optimization tips will flag bloat directly.