From the source
- Google is previewing Tool Retrieval in the Gemini Interactions API, letting agents keep large tool libraries on standby and load full parameter schemas only when a task needs them.
- Google says sending every full schema inflates input-token counts, adds latency and reduces tool-selection accuracy as the candidate list grows, so deferred tools initially contribute only their name and description.
- In the default server-side mode, developers add a tool_search tool and set defer_loading: true on selected function tools; Gemini performs the search, so there is no index to build or host, and setting defer_loading on a remote MCP server defers every tool it exposes.
- Client-side retrieval moves search, ranking and filtering to the developer's application, adding one round trip per search but allowing tools to be introduced mid-conversation; the two modes cannot be combined in one request.
- Tool Retrieval carries no separate charge, but the setup guide reports early testing produced about 26 tokens per deferred tool, and every loaded schema stays billable as input on later calls.
- The guide's illustrative example puts 200 tools of 400 tokens each with two loaded at 164,000 input tokens statically versus about 23,650 with retrieval, while a 10-tool library with a 30,000-token context rises from 68,000 to about 92,830 tokens.
- The guide lists Gemini 4 Argon, Gemini 3.8 Flash and Gemini 3.6 Flash as supported, returns a 400 error for defer_loading on unsupported models, and warns the preview surface may change before general availability.