# Google adds Tool Retrieval to the Gemini API in preview

published: 2026-11-19T19:05:22.861Z
topic: infrastructure
source: https://blog.google/innovation-and-ai/technology/developers-tools/gemini-api-tool-retrieval/
url: https://aiagentnews.io/stories/google-adds-tool-retrieval-to-the-gemini-api-in-preview

Google has launched Tool Retrieval in preview in the Gemini Interactions API, letting agents load tool parameter schemas on demand rather than sending the full tool library with every prompt. The feature is aimed at lowering input token use and keeping tool selection accurate as the number of tools grows.

Tool Retrieval lets an agent keep its full tool library on standby. Developers add a tool_search tool to a request and set defer_loading to true on the tools to hold back. The model still sees each deferred tool's name and description, but the parameter schemas stay out of context until the model asks for a tool by name.

When that happens, the API loads the full schema and the model emits the function call within the same request. Each step appears in the interaction's steps list, which makes it easier to check which tools the model considered. Gemini runs the search on the server, so there is no index to build or host.

Developers who maintain their own tool registry can instead declare tool_search with execution set to client and run the lookup in their own app, which adds one round trip per search. Remote MCP servers are also supported: setting defer_loading on an mcp_server tool loads each exposed tool on demand. Tool Retrieval has no separate charge, and it works with Gemini 3.6 and later models, including Gemini 4 models.

## Why it matters if you build agents

Agents with dozens or hundreds of tools can keep frequently used tools loaded up front and defer the rest, which may cut input tokens and latency.

Savings depend on usage: each schema the model loads is billed as input for the rest of the conversation, so developers should measure the effect on their own traffic.

Because deferred schemas stay out of the start of the prompt, prefix caching should remain effective across turns.
