MCP Tool Search - Claude Code

MCP tool search solves the "my MCP server has 80 tools" problem. Tools are loaded on demand instead of all at once.

What it does

When a server has tool search enabled, Claude Code sees a searchable index rather than every tool loaded into context. The model queries the index when it needs a tool, loads the matching tool's schema, and calls it. You keep the expressive power of large tool suites without paying tokens for every tool, every turn.

When to use it

• Any MCP server exposing more than a handful of tools.
• Multi-server setups where combined tool count bloats context.
• Cost-sensitive workflows where tool schemas were eating the budget.
• Large internal platforms with hundreds of operations.

Gotchas

• Tool search adds a small latency for the first call to each tool.
• Search queries affect tool selection quality. Servers should provide good descriptions.
• Not every MCP server supports deferred loading yet - check the server docs.

Official docs: https://code.claude.com/docs/en/mcp.md#scale-with-mcp-tool-search