Google adds Tool Retrieval preview to the Gemini API

Agents can keep large tool libraries on standby and fetch parameter schemas only when needed, which Google says cuts input tokens and helps tool-selection accuracy.

AIPublished
Image published by Google Blog with “Introducing Tool Retrieval in the Gemini API”
Image: Google Blog · Source
What happened
On 19 November Google launched Tool Retrieval in preview in the Gemini Interactions API.
What’s new
Developers can defer tool parameter schemas so the model loads them by name only when a task requires them.
Why it matters
Google says sending every schema with every request inflates token counts, adds latency and reduces tool-selection accuracy for agents with many tools.

Google on 19 November released Tool Retrieval, a preview feature in the Gemini Interactions API aimed at agents connected to dozens or hundreds of tools. According to Google, including every tool's full parameter schema in each request raises input token counts, slows responses and makes the model worse at picking the right tool as the list grows.

With Tool Retrieval, developers add a tool_search tool to a request and mark tools they want held back with a defer_loading flag. The model still receives each deferred tool's name and description, but not its parameters. When it needs one, it requests it by name, the API supplies the schema, and the model makes the function call, all within a single request. Google runs the search server-side, so no index has to be built. Setting the flag on a remote MCP server tool defers every tool that server exposes.

Teams with their own tool registry or permission filters can instead run the search in their own app, which Google says costs one extra round trip per search but allows new tools to be exposed mid-conversation.

There is no extra charge; standard token rates apply. Google notes that switching on search brings a modest, fixed number of extra input tokens, and loaded schemas are billed as input, so savings depend on how many schemas go unused; it advises developers to measure against their own traffic. The feature supports Gemini 3.6 and later, including all Gemini 4 models.