Open Brain
GoogleOfficial ReleaseOfficial Source

Introducing Tool Retrieval in the Gemini API

Gemini Tool Retrieval defers full tool schemas until an agent needs them, keeping large tool libraries out of most requests. Measure token use on your traffic, since loaded schemas remain billable in the conversation.

Google · Nov 19, 2026
Open Source Open MarkdownOpen JSON
Source Summary

**Tool Retrieval** is in preview in the Gemini Interactions API. Deferred tools keep their names and descriptions visible, while their parameter schemas load only after the model requests a specific tool; the search and call happen in one API request.

Practical Implication

For agents with large tool libraries, leave frequent tools loaded and defer the long tail with `tool_search` and `defer_loading: true`. Remote MCP tools can also be loaded on demand, and deferred schemas keep the prompt prefix stable for caching.

Agent-Ready Context
**Tool Retrieval** is in preview in the Gemini Interactions API. Deferred tools keep their names and descriptions visible, while their parameter schemas load only after the model requests a specific tool; the search and call happen in one API request.

For agents with large tool libraries, leave frequent tools loaded and defer the long tail with `tool_search` and `defer_loading: true`. Remote MCP tools can also be loaded on demand, and deferred schemas keep the prompt prefix stable for caching.

There is no separate feature charge, but search adds a small fixed input-token block and loaded schemas are billed as input for the rest of the conversation. Savings depend on how many schemas remain unused, so measure against your own traffic.
Connected Context · Feed7 Judgment

Tool Retrieval adds an API-level way to keep large tool libraries discoverable while loading schemas only when needed, including remote MCP tools. This makes deferred schema loading a practical context-management option alongside compressing or externalizing context; its value still depends on how many tools stay unused and on the cost of the added search tokens.

Context Map
contextcoding#tool-use#context-engineering#context-caching
Uncertainty
There is no separate feature charge, but search adds a small fixed input-token block and loaded schemas are billed as input for the rest of the conversation. Savings depend on how many schemas remain unused, so measure against your own traffic.