On November 19, Google introduced Tool Retrieval in preview in the Gemini API. When an agent is wired to dozens or hundreds of tools, developers no longer have to stuff every tool's parameter schema into each request. The model first sees only tool names and descriptions, fetches the full definition by name when it needs one, and then issues the function call within the same request.
For now the feature is available only in the Interactions API, on Gemini 3.6 and later, including all Gemini 4 models. Google's rationale in its blog post is straightforward: with many tools, sending every full schema on every request drives up input tokens and latency, and the more candidate tools there are, the less likely the model is to pick the right one.

Two ways to retrieve
The default is server-side retrieval. Developers add {"type": "tool_search"} to the tools list and mark rarely used tools with defer_loading, while frequently used tools are declared directly as before. The search runs on Google's servers, so there is no index to build or host. For debugging, interaction.steps exposes three step types, tool_search_call, tool_search_result and function_call, which helps pin down whether a tool's name or its description is the unclear part.
The alternative is client-side retrieval: declare tool_search with "execution": "client", and when the model issues a search request, the application queries its own tool registry and returns the matching schemas as the function result. Each search costs an extra round trip, but in return tools don't have to be declared up front, new tools can be added mid-conversation, results can be filtered by user permissions, and the app can plug in its own vector index.
Remote MCP servers are covered too: set defer_loading: true on an mcp_server tool and every tool that server exposes switches to on-demand loading. Google also notes that deferred schemas are not placed at the start of the prompt; schemas loaded later are appended further down the conversation, so the prefix cache across turns is not invalidated.
"Tool Retrieval has no separate charge. Input and output tokens are billed at the standard rates for the model you call."
Whether it saves money depends on the size of the tool library
Turning on tool_search adds a small fixed block of input tokens; the blog doesn't say how large. Every schema the model loads is billed as input for the rest of the conversation, just like a tool declared up front. Google's position is that tokens are saved only when more schemas stay unloaded than get loaded, which is usually the case with large tool libraries, and it advises developers to test with their own real traffic. Conversely, for apps with few tools, or ones that use most of their tools on every turn, enabling it may not pay off.
Same idea, with Anthropic first out of the gate
Claude's developer platform introduced this approach in November last year. Anthropic calls it the Tool Search Tool, uses the same defer_loading parameter name, and likewise brings MCP servers into on-demand loading. The difference lies in the evidence offered: Anthropic published the token savings and tool-selection accuracy from its internal tests, while Google this time describes only the mechanism and billing, with no results data at all.
For teams that already build agents on the Claude or OpenAI APIs, the main hurdle this time is the API itself. Tool retrieval works only in the Interactions API, so Gemini projects still on the older API must migrate before they can switch it on. Google has also not yet published quotas or rate limits for after the preview ends.
Sources: Google developer blog, CocoLoop, Google AI Studio tool retrieval documentation, Anthropic developer documentation; we checked the supported model versions, the two retrieval modes, MCP deferred loading and the billing rules.