Google has introduced Tool Retrieval in preview for the Gemini API, intended to help agents manage large libraries of tools without sending every tool schema on each request. The announcement was posted by Lucia Loher, Product Manager, Gemini API, and Nora Kassner, Research Scientist, Google DeepMind.
How Tool Retrieval works
Tool Retrieval keeps a full tool library on standby while presenting only names and descriptions to the model. Parameter schemas are withheld until the model requests a specific tool by name. When required, the API loads that tool’s schema and the model can perform the function call within the same request. The Interactions API records these steps—tool_search_call, tool_search_result and function_call—in interaction.steps to aid inspection and debugging.
Developers enable the feature by adding a tool_search tool to requests and setting defer_loading: true on tools to be deferred. The model still sees tool names and descriptions, but schemas are added later only when invoked.
Client-side retrieval, caching and MCP servers
Tool Retrieval supports client-side retrieval: declare the tool_search with “execution”: “client” and let the application run lookups or permission filters. This adds one round trip per search and allows exposing new tools mid-conversation. The feature also integrates with prefix caching: deferred schemas remain out of the conversation prefix, and schemas loaded later do not alter cached prefixes between turns. For remote MCP servers, setting defer_loading: true on an mcp_server tool causes the server’s tools to be loaded on demand instead of listed individually.
According to Google, there is no separate charge for Tool Retrieval. Input and output tokens are billed at the model’s standard rates; enabling tool_search adds a small fixed input token cost and each loaded schema is billed as input for subsequent turns. Google recommends measuring token savings against actual traffic, since savings occur when many declared schemas are never loaded.
Tool Retrieval is available in preview in the Interactions API and works with Gemini 3.6 and later, including all Gemini 4 models. It can be enabled by adding a tool_search tool and setting defer_loading: true on deferred tools.
Original source: Google AI