Google has introduced Tool Retrieval in preview for the Gemini Interactions API, targeting agents connected to dozens or hundreds of functions. Instead of placing every complete parameter schema in every prompt, developers can defer less-used tools and let Gemini request their details only when needed.
The model still sees the name and description of each deferred function. When it chooses one, the API searches for that tool, loads the full schema and produces a function call inside the same request. Frequently used tools can remain fully loaded, and remote Model Context Protocol servers can also expose their tools on demand.
Developers with a private registry or per-user permission filter can run retrieval on the client side. That path adds a round trip because the application performs the search and returns matching schemas, but it allows tools to appear later when a user connects a new integration.
Tool Retrieval has no separate fee and works with Gemini 3.6 and later models in the Interactions API. The search adds a fixed block of input tokens, while loaded schemas are billed as normal input. Savings therefore depend on how many schemas remain unused. Teams should measure both token cost and selection accuracy on real traffic rather than assume a large tool library will automatically become cheaper.