Skip to content
worth noting New models

Google launches Tool Retrieval in the Gemini API for deferring tool schemas

clearly official source

Google has launched, in preview, Tool Retrieval in the Gemini API — a feature that defers the loading of tool parameter schemas until the agent actually needs them, reducing input token consumption for agents with many tools.

Google has launched, in preview, a Tool Retrieval feature in the Gemini API, available as part of the Interactions API. The feature addresses the problem where an agent connected to dozens or hundreds of tools must send their full parameter schemas with every request — this increases the number of input tokens, slows down response time, and, according to Google, reduces the accuracy of selecting the correct tool as the number of candidates grows.

It is enabled by adding the tool_search tool to the request and setting the defer_loading: true flag on tools that should remain deferred. The model sees the name and description of each deferred tool, so it knows of its existence, but the full schema is loaded only at the moment the model needs it — both the lookup and the schema loading happen within a single API request, followed by the function call. The search runs server-side, so the developer doesn't need to build or host their own index; alternatively, one can choose "execution": "client" and perform the search in one's own tool registry, which also allows offering new tools during the conversation. The feature is compatible with deferred-priority prefix caching, because deferred schemas stay outside the beginning of the prompt and don't disrupt the shared cache between conversation turns. It also supports remote MCP servers — by setting defer_loading on mcp_server, all tools of that server are loaded on demand.

According to Google, Tool Retrieval is not billed separately; standard rates for input and output tokens of the given model apply. Enabling tool_search adds a small fixed block of input tokens, and each loaded schema is then billed as input for the rest of the conversation, the same as a tool declared upfront — the savings materialize when the number of unloaded schemas outweighs the loaded ones, which, according to the company, is the usual case for large tool libraries. The feature is available for Gemini 3.6 models and newer, including all Gemini 4 models, as part of the Interactions API.

What changed

Why it matters

Developers of agentic systems with a large number of tools (dozens to hundreds) can, thanks to deferred schema loading, reduce the number of input tokens sent with every request, speed up response times, and improve the accuracy of selecting the correct tool — without having to build their own search index. Moreover, the feature doesn't disrupt prefix caching, so the savings from caching are preserved even when using a large number of tools.

Two audiences, two different impacts

What this means

01

For individuals

A developer building agents with many tools via the Gemini API can reduce input token consumption and improve tool-selection accuracy by deferring the loading of parameter schemas, instead of sending them all at once with every request.

What to do Try the tool_search tool with the defer_loading flag on agents with a large number of functions and compare token consumption and tool-selection accuracy.
More practical updates →
02

For a business

Companies building agentic systems with a large number of integrated tools or MCP servers can reduce input token costs while preserving prefix caching functionality, without having to build their own indexing — according to Google, the feature is not billed separately; standard rates for input and output tokens apply.

Development
What to decide If a company runs agents with dozens or hundreds of tools via the Gemini API, verify the benefit of the defer_loading feature in your own deployment and measure token savings.
More business impacts →
agents API development function-calling Gemini API Tool Retrieval

Check the original

Event sources

clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.

1
Google AI blog (The Keyword — AI) primary source · first detected Introducing Tool Retrieval in the Gemini API