The development
What happened
Google introduced Tool Retrieval on 19 November, in preview in the Gemini Interactions API, to let AI agents work with large tool libraries without receiving every parameter schema upfront. The model still sees deferred tools’ names and descriptions; it requests the schemas it needs before emitting function calls. Google says this can reduce input tokens and latency and help tool selection as libraries grow. Those are workload-dependent benefits, not a published universal savings or accuracy guarantee. Developers can keep frequently used tools loaded, defer the rest, or run retrieval through their own registry.
Key facts
Enable server-side retrieval with a tool_search tool and defer_loading on selected functions. Names and descriptions remain visible; only parameter schemas are deferred.
The model requests tools by name, the API loads their schemas, then the model emits function calls in the same request. Application code still executes the functions and returns results.
Server-side retrieval requires no separately hosted index and supports deferred remote MCP servers. Client-side retrieval can use an app’s registry or permission filter, adding one round trip per search.
Google says deferred schemas can preserve a stable cached prefix: loaded schemas are added later in the conversation. Keeping common tools undeferred remains an option.
There is no separate retrieval fee, but standard input/output rates apply. tool_search adds fixed input tokens, and loaded schemas remain billable input through the conversation; savings depend on what is never loaded.
The preview supports Gemini 3.6 and later models, including all Gemini 4 models, in the Interactions API. This is not a claim of availability in every Gemini endpoint or app.
The context
Why it matters
TideReach assessment: the useful tradeoff is between carrying all schemas and paying retrieval overhead for a smaller working set. Teams should measure their own tool mix and cache behavior rather than assume automatic savings. Keeping tool discovery separate from application execution also leaves validation and user-permission checks with the application; retrieving a schema is not approval to perform its action.