Skip to content
The Daily Circuit
AI News

Google Introduces Tool Retrieval Feature in Gemini API for Large Tool Libraries

Google has launched Tool Retrieval in the Gemini API preview, allowing AI agents to load parameter schemas on demand to reduce input tokens and latency.

By The Daily Circuit AI Desk (automated, AI-generated)

2 min read

Abstract digital illustration representing AI tool retrieval and token optimization

AI-generated image (flux-1-schnell), not a photograph of the events described.

Key facts

  • Tool Retrieval is available in preview in the Gemini Interactions API.[1]
  • The feature works by loading parameter schemas on demand when deferred with defer_loading set to true.[1]
  • Tool Retrieval is compatible with Gemini 3.6 and later models, as well as all Gemini 4 models.[1]
  • There is no separate charge for Tool Retrieval; standard model rates apply for input and output tokens.[1]
  • Tool Retrieval supports remote MCP servers by loading exposed tools on demand.[1]

Google AI has announced the launch of Tool Retrieval in the Gemini API, currently available in preview within the Interactions API [1]. The feature is designed to help developers manage extensive software tool collections more efficiently by loading tool parameter schemas on demand rather than sending full tool libraries in every prompt [1].

Efficient Management of Large Tool Libraries

When AI agents connect to dozens or hundreds of tools, sending every full parameter schema on every request increases input token counts, introduces latency, and can reduce tool-selection accuracy [1]. Tool Retrieval allows an agent to keep its entire tool library on standby while loading parameter schemas only when required for a specific task [1].

To enable the feature, developers add a tool_search tool to their request and set defer_loading to true on the tools they wish to hold back [1]. Although the parameter schemas are kept out of the immediate context, the model still retains visibility of the name and description of every deferred tool so it knows what capabilities are available [1]. When a task requires a specific tool, the model requests it by name, the API loads the full schema, and the model executes the function call within the same request [1].

Integration and Configuration

Google notes that the feature requires no separate index to be built or hosted, as Gemini handles the search on the server side [1]. Code modifications are limited to adding the new tool and setting the flag on desired deferred tools [1]. Developers can choose to keep their most frequently used tools loaded upfront while deferring the long tail of less common tools [1].

The feature also supports remote Model Context Protocol (MCP) servers in the Gemini API [1]. By setting defer_loading to true on an mcp_server tool, every tool exposed by that server is loaded on demand, eliminating the need to list individual tools one by one [1]. For developers maintaining custom tool registries, embedding indexes, or per-user permission filters, client-side retrieval can be declared by setting tool_search execution to client [1].

Compatibility and Pricing

Tool Retrieval is compatible with Gemini 3.6 and later models, alongside all Gemini 4 models, within the Interactions API [1]. There is no separate charge for Tool Retrieval, and input and output tokens are billed at standard model rates [1]. Enabling tool_search adds a small, fixed block of input tokens, and any schema loaded by the model is billed as input for the remainder of the conversation [1].

Context

As AI agents scale and integrate with dozens or hundreds of external software tools, managing prompt context windows efficiently has become critical. Sending complete parameter schemas for every available tool on every API call consumes valuable input tokens and can impact overall model performance.

Why it matters

For developers and enterprises building complex AI agents in India and globally, optimizing token usage directly impacts operational costs and application responsiveness. Tool Retrieval offers a structured way to scale agent capabilities without a linear increase in input token overhead.

What could happen next

AI-generated analysis based on the sources below. Not a forecast or recommendation.

The introduction of server-side and client-side tool retrieval could lower operational costs for large-scale enterprise deployments of AI agents, making more complex multi-tool integrations economically viable.

Risks and open questions

Developers must measure token savings against their specific traffic patterns, as loading schemas mid-conversation still incurs ongoing input token billing for the remainder of the interaction.

Sources

  1. 1.Google AI (primary source): Introducing Tool Retrieval in the Gemini API, 19 November 2026

How this article was produced. This story was researched and written by an automated system using Google’s Gemini models, working only from the sources listed above. Automated checks verified that every citation, quote and figure matches those sources before publication. It was not written by a human reporter. Editorial policy · Report an error

  • Gemini API
  • Google AI
  • Artificial Intelligence
  • Developer Tools