Google has launched Tool Retrieval in the Gemini API, now available in preview in the Gemini Interactions API. Instead of sending an agent's full tool library with every request, developers can keep tools on standby and load each one's parameter schema only when the current task needs it.

The feature is aimed at a scaling problem: when an agent connects to dozens or hundreds of tools, sending every full parameter schema on every request inflates input token counts, adds latency, and makes tool selection less accurate as the candidate list grows.

To turn it on, developers add a tool_search tool to the request and set defer_loading to true on the tools they want to hold back. The model still sees every tool's name and description, so it knows what is available, but the parameter schemas stay out of context until the model asks for a tool by name. The API then loads the full schema and the model calls the function, all inside the same request, with each step visible for debugging.

There is no index to build or host, since Gemini runs the search on the server, and the only code changes are the new tool and a flag on the tools to defer. Developers can keep their most-used tools loaded up front and defer the long tail, and deferred schemas work with prefix caching because they stay out of the start of the prompt.

Tool Retrieval also works with remote MCP servers in the Gemini API: setting defer_loading to true on an mcp_server tool loads every tool that server exposes on demand. A client-side option is available too, for teams that run their own tool registry, embedding index, or per-user permission filter.

There is no separate charge. Input and output tokens are billed at the standard rates for the model, with a small fixed block of input tokens for tool_search and each loaded schema billed as input for the rest of the conversation. Tool Retrieval works with Gemini 3.6 and later models, including all Gemini 4 models, in the Interactions API.