Defer Loading and Tool Search
Definition
defer_loading withholds a tool's description from the initial model context; the server-side Tool search tool then fetches definitions on demand, keeping large tool catalogs cheap.
Key points
defer_loading: true(defaultfalse) — when true, the tool's description is not sent to the model initially; it is discovered via the Tool search tool.- Used with large tool sets to save context tokens.
- Available both as a per-tool field on custom tool definitions and inside
mcp_toolsetper-toolconfigs. - The Tool search tool is a server tool (runs on Anthropic infra; no handler code).
Why it matters for the exam
- Tested as the context-efficiency answer when a server exposes dozens/hundreds of tools.
- Pairs conceptually with Adaptive Thinking and Effort and caching as a context-budget lever.
Common gotchas
defer_loadingis a token/context optimization, not an access control — use MCPToolset Allowlist and Denylist to actually restrict tools.- Deferred tools cannot be called until the Tool search tool has surfaced their schema.
See also
Sources
Referenced by
Practice questions optional · AI
Generate fresh practice questions about this concept with AI. These are not vault-verified.