Google is launching a new "Tool Retrieval" feature in the Gemini Interactions API, now available in preview for developers building agent-based applications. The update aims to reduce input token costs and latency when agents interact with large libraries of software tools.

The feature addresses a scaling bottleneck in complex AI agents. When an agent connects to dozens or hundreds of tools, sending every full parameter schema on every request inflates input token counts and adds latency, while lowering tool-selection accuracy as the candidate list grows. Tool Retrieval allows the model to keep its full tool library on standby, loading specific parameter schemas only when the current task requires them.

According to Google’s announcement, the system works by adding a tool_search tool to the request and setting defer_loading: true on the tools to be held back. The model still sees the name and description of every deferred tool, enabling it to know what is available without consuming context space for the full schema. When the model identifies a need for a specific tool, it requests it by name within the same API call; the API then loads the full schema, and the model emits the function call. All three steps appear in interaction.steps, allowing developers to inspect which tools the model considered during debugging.

The implementation supports both server-side and client-side retrieval. For server-side operations, Gemini runs the search internally, requiring no index hosting from the developer. This approach also integrates with prefix caching, as deferred schemas remain outside the cached prompt prefix. For client-side scenarios, developers can declare the tool_search with "execution": "client" to use their own tool registry or embedding index, though this adds one round trip per search.

Tool Retrieval is supported by Gemini 3.6 and later models, including all Gemini 4 models, such as Gemini 4 Neon Introducing Tool Retrieval in the Gemini API1. The feature incurs no separate charge; input and output tokens are billed at standard rates. Loading a schema bills as input for the remainder of the conversation. Developers save tokens when the count of never-loaded schemas outweighs those loaded, a scenario Google notes is typical for large libraries.

Automatica Press reported in November that Google DeepMind recently introduced Gemini 4 Neon for code and science tasks, establishing the model lineage now supporting this API update Google DeepMind Launches Gemini 4 Neon AI Model for Code and Science2.

The announcement includes no independent test results verifying the claimed reductions in latency or token usage. Pricing details confirm standard billing but do not specify exact token savings percentages for specific use cases.