# Gemini API adds Tool Retrieval to load agent tools only when needed

> The preview feature keeps tool schemas out of the prompt until the model asks for one, to cut input tokens for agents with many tools.

URL: https://claudegpt.si/news/gemini-api-tool-retrieval  
Published: 2026-11-19T18:57:09.890Z  
Author: ClaudeGPT Si Newsroom  
Section: AI Tools  
Topics: Google, Gemini API, AI agents, MCP, Developer tools  
Publisher: ClaudeGPT Si

## Key takeaways

- Google added Tool Retrieval to the Gemini API's Interactions API in preview on November 19, 2026.
- Deferred tools show only a name and description until the model requests the full schema.
- It works with Gemini 3.6 and later models and has no separate fee beyond normal token prices.

Google added Tool Retrieval to the Gemini API on November 19, in preview in its Interactions API. It lets an agent keep a large library of tools on standby and load each tool's full definition only when a task calls for it, according to [Google's announcement](https://blog.google/innovation-and-ai/technology/developers-tools/gemini-api-tool-retrieval/).

The problem it targets is familiar to anyone who has wired dozens of functions or MCP servers into one agent. Every tool's parameter schema normally rides along with every request. Google says that "inflates input token counts, adds latency, and lowers tool-selection accuracy." With Tool Retrieval, the model sees only each deferred tool's name and description. When it needs one, it asks for it by name, the API loads the schema, and the model makes the call, all inside one request. Each step shows up in the response as a `tool_search_call`, a `tool_search_result` and a `function_call`.

Turning it on takes two changes: add a `{"type": "tool_search"}` tool to the request and set `defer_loading` on the functions to hold back. The search runs on Google's servers, so there is no index to build. Teams with their own tool registry or permission rules can run the search themselves instead, at the cost of one extra round trip per lookup. Remote MCP servers work too: `defer_loading` on an MCP server loads its tools on demand.

It works with Gemini 3.6 and later models, including all Gemini 4 models. There is no separate fee. Enabling it adds a small fixed block of input tokens, and every schema the model loads is billed as input for the rest of the conversation. Google gives no numbers for the savings and tells developers to measure them on their own traffic: an agent that ends up loading most of its tools may save little. The [setup guide](https://aistudio.google.com/docs/tool-retrieval) is in the AI Studio docs.

For what the newest model costs to run, see [Google releases Gemini 4 Neon](https://claudegpt.si/news/gemini-4-neon-release).

## Sources

- [Introducing Tool Retrieval in the Gemini API](https://blog.google/innovation-and-ai/technology/developers-tools/gemini-api-tool-retrieval/), Google
- [Tool retrieval](https://aistudio.google.com/docs/tool-retrieval), Google
