Google has launched Tool Retrieval in the Gemini API, in preview in the Interactions API. The feature lets AI agents keep a library of dozens or even hundreds of tools on standby and load each tool's parameter schema only when the task requires it — instead of sending every full definition on every request.
According to the announcement, published on November 19, 2026 by Lucia Loher, product manager for the Gemini API, and Nora Kassner, research scientist at Google DeepMind, sending every schema on each call inflates input token counts, adds latency, and lowers tool-selection accuracy as the list grows.
How it works
To turn it on, the developer adds a tool_search tool to the request and marks the tools to hold back with defer_loading: true. The model still sees the name and description of every deferred tool, so it knows what is available, but the parameter schemas stay out of context until requested. When the model needs a tool, it asks for it by name, the API loads the full schema, and the model emits the function call — all inside the same request.
Google's example is a customer-service agent: a customer reports that order A-1042 arrived damaged and asks for a 40-dollar refund. The model searches for the issue_refund function, the API loads its schema, and the call is made. All three steps — tool_search_call, tool_search_result and function_call — appear in interaction.steps, making it easy to inspect which tools the model considered.
There is no index to build or host: the search runs on Gemini's server. The most-used tools can stay loaded up front, with only the long tail deferred. The feature also works with prefix caching: deferred schemas stay out of the start of the prompt, so the cached prefix stays the same from call to call.
Pricing, models and MCP
Tool Retrieval has no separate charge: input and output tokens are billed at the standard rates for the model called. Turning on tool_search adds a small, fixed block of input tokens, and each schema the model loads is billed as input for the rest of the conversation. Savings appear when the schemas never loaded outweigh the ones that are — the usual case for large libraries, according to Google.
The feature works with Gemini 3.6 and later models, including all Gemini 4 models, and only in the Interactions API. It also works with remote MCP servers: setting defer_loading: true on an mcp_server tool loads every tool that server exposes on demand. There is also a client-side retrieval variant for developers who maintain their own tool registry — in that mode, each search adds one round trip, and new tools can be exposed mid-conversation.
The idea is not new: Anthropic launched a tool-search feature in November 2025 and OpenAI in March 2026, with a similar design. Google gave no measured figures for token savings, latency or tool-selection accuracy, and no date for general availability — as a preview, the interface may change before a stable release.
The feature arrives the same week Google opened the Genie 4 API to developers: Update: Google opens the Genie 4 API to developers.
Image credit: Google / Wikimedia Commons
Sources
- Google — Gemini API tool retrieval for efficient agent scaling
- TalkAndroid — Gemini API Tool Retrieval Loads Agent Tools On Demand
- PPC Land — Google shrinks idle tools in Gemini agents to about 26 tokens each
Transparency: this article was produced with the help of artificial intelligence, by cross-checking public posts on X and sources on the internet. Please consult the original sources to confirm the information.
By Muse