本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://x.com/trq212/status/2011523109871108570
Summary
Claude Code ships MCP Tool Search — a lazy loading mechanism that detects when MCP tool descriptions would consume more than 10% of context and switches to dynamic search-based loading. Users were reporting setups with 7+ servers consuming 67k+ tokens. The feature resolves the most-requested GitHub issue for Claude Code. Server authors should focus on the “server instructions” field to help Claude know when to search for their tools.
Claude Code 推出 MCP Tool Search：當 MCP 工具描述超過 10% 的 context 時，自動切換為搜尋式按需載入。這解決了最多人請求的 GitHub Issue。有用戶回報 7 個以上 MCP server 消耗 67k+ tokens。server 作者應充實 “server instructions” 欄位以幫助 Claude 判斷何時搜尋工具。
Key Points
- Triggers when MCP tool descriptions > 10% of context window
- Tools loaded via search rather than preloaded into every session
- “Server instructions” field becomes important as a hint for when to search
- MCP client developers should implement ToolSearchTool (docs linked)
- Programmatic tool calling (composing MCP tools via code) is future work
Insights
This is an architectural response to the context window economics problem for plugin-heavy setups. The 10% threshold is a pragmatic heuristic. The shift from “preload everything” to “search and load” mirrors how good RAG systems work — lazy retrieval beats eager loading when the full tool set exceeds useful context budget.
Connections
Raw Excerpt
“Users were documenting setups with 7+ servers consuming 67k+ tokens. Tool Search allows Claude Code to dynamically load tools into context when MCP tools would otherwise take up a lot of context.”