Google on Nov. 19 launched a preview of Tool Retrieval in the Gemini API, a feature meant to help AI agents work with large numbers of software tools. The company said it is available in the Gemini Interactions API.

According to Google, an agent connected to dozens or hundreds of tools has to send each tool’s full parameter schema with every request. A parameter schema is the detailed description of the inputs a tool accepts. Google said this raises input token counts, adds delay and lowers the accuracy of tool selection as the list grows. Tool Retrieval lets an agent keep its full library on standby and load those details only when a task calls for them, the company said.

Developers turn the feature on by adding a tool called tool_search to a request and setting a flag, defer_loading, to true on the tools to hold back. Google said the model still sees each deferred tool’s name and description. When the model needs a tool, it requests that tool by its name. The API then loads the complete schema and the model calls the function, with every step happening inside a single request. Google said the steps appear in the interaction record, so developers can check which tools the model considered.

Google said there is no index to build or host, because Gemini runs the search on its servers. Developers can keep frequently used tools loaded and defer the rest. The company said the feature also works with remote MCP servers: setting the flag on a server’s tool entry loads every tool that server offers on demand.

Developers who keep their own tool registry or per-user permission filter can run the search themselves by declaring the search tool with client-side execution. Google said this adds one round trip per search, because the developer’s app performs the lookup. It also lets developers expose new tools in the middle of a conversation.

Google said the feature is compatible with prefix caching. It said deferred schemas stay out of the start of the prompt, so the cached prefix stays the same from turn to turn.

On pricing, Google said there is no separate fee for Tool Retrieval. Input and output tokens are charged at the usual rates for whichever model a developer calls. Switching on the search tool adds a small block of input tokens that stays the same size, and every schema the model loads counts as input for the remainder of the conversation. Google said tokens are saved when the schemas never loaded outweigh those that are, which it called the usual case for large libraries. It advised developers to measure the effect against their own traffic.

In the Interactions API, Tool Retrieval is supported on Gemini 3.6 and newer models, a group that includes every Gemini 4 model. The post was written by Lucia Loher, a product manager for the Gemini API, and Nora Kassner, a research scientist at Google DeepMind.