- 4 min read
- 8 key claims
- 1 source cited
- For Developers, AI makers, Businesses
Overview
Tool Retrieval is a preview feature in the Gemini Interactions API for agents that connect to many software tools. It is intended for large tool libraries in which sending every parameter schema with every request adds context and processing overhead. Developers enable it by adding a tool_search tool and setting defer_loading: true on tools they want to hold back [1].
With deferred loading, the model retains the name and description of each deferred tool, while its full parameter schema remains outside the initial context. When the model needs a tool, it requests it by name, the API loads the schema, and the model can issue the function call within the same request [1].
The feature works with Gemini 3.6 and later models. It can also be applied to an mcp_server tool so that every tool exposed by a remote MCP server is loaded on demand, without declaring those tools individually [1].
Why it matters
On-demand schema loading changes how a developer organizes an agent with a broad tool catalog. Frequently used tools can remain loaded, while less common tools stay deferred until the model requests them. That makes it possible to keep a larger set of tools available without placing every schema in each prompt [1].
The approach may be most useful where schema volume, rather than tool count alone, drives context overhead, because deferred tools still leave their names and descriptions in context. Remote MCP support extends the same pattern to tools that are not individually listed by the developer [1].
Google says deferred schemas leave the start of the prompt unchanged, which allows a cached prefix to remain stable across turns. Schemas loaded later are appended to the conversation [1].
Opportunities
- Keep a core set eager. Developers can leave common tools loaded and defer the long tail, rather than deciding whether to load the entire library or restrict the catalog [1].
- Connect large remote MCP servers. Setting
defer_loading: trueon an MCP server tool makes its exposed tools available for on-demand loading without a per-tool declaration [1]. - Reuse an existing registry. Client-side retrieval fits applications that already operate a tool registry, embedding index or permission-filtered catalog, since the application can perform the lookup and return schemas [1].
- Add tools during a conversation. In the client-side setup, tools do not need to be declared upfront, which supports exposing new tools when a user connects an integration [1].
- Inspect the selection process. The interaction steps include a tool-search call, a search result and a function call, giving developers visibility into which tools the model considered when debugging names or descriptions [1].
Risks
Client-side retrieval introduces one additional network round trip for each search, which can offset latency gains in applications that supply their own lookup service [1].
The feature remains a preview in the Gemini Interactions API, so adopting it ties an agent's tool-loading behavior to that API and its supported model range [1].
Deferred tools still expose their names and descriptions to the model. Google does not address how those descriptions affect tool choice, or whether sensitive metadata should be included in them [1].
The documentation does not discuss security or privacy implications of loading schemas on demand, including how schemas returned from a client-side registry or remote server are handled [1].
Context
Ordinary function calling requires the API to send the tool library with each request. As that library grows, the parameter schemas add to the input token count for every turn. Google describes this as a problem particularly relevant to agents connected to dozens or hundreds of tools [1].
Tool Retrieval separates a tool's summary from its callable details. Names and descriptions remain visible for tool selection, but full parameter schemas are fetched only when required. The model-side search runs on Google's servers, so developers do not need to build or host an index for this approach [1].
For applications that already maintain their own registry, embedding index or per-user permission filter, Google also documents client-side retrieval. Developers set "execution": "client" on the search tool, and the application returns matching schemas. This adds one network round trip per search because the lookup runs in the developer's application rather than on Google's servers [1].
Known, claimed, inferred
Google presents Tool Retrieval as a way to scale agents toward large tool catalogs without permanently reserving context for every parameter schema. The mechanism is specific: defer the schema, keep the description, retrieve the detail on request and continue with ordinary function calling inside the same interaction [1].
Two retrieval paths are documented. Server-side search requires no separate index and keeps the lookup inside the API request. Client-side search gives developers control over an existing registry, but moves the lookup into their application and introduces a round trip [1].
The stated benefits are therefore coupled to a particular retrieval design. Token savings depend on how much of each deferred tool's metadata is schema rather than name or description, and latency outcomes depend on where the search runs. The available announcement does not quantify either effect [1].
Limitations & uncertainties
Google's announcement provides no measured figures for input-token savings, latency reduction or tool-selection accuracy. There are no independent benchmarks or third-party test results, and no stated limits on the number of deferred tools or measurements of behavior at scale [1].
Beyond the absence of a separate charge, the announcement does not set out further pricing details. Input and output tokens are billed at standard rates [1].
Evidence & sources
The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.
- Primary-source or vendor statementTool Retrieval is available in preview in the Gemini Interactions API.[1]
- Primary-source or vendor statementDevelopers enable it by adding a `tool_search` tool and setting `defer_loading: true` on tools to defer.[1]
- Primary-source or vendor statementDeferred tools keep their names and descriptions in context while their parameter schemas are loaded only when the model requests them.[1]
- Primary-source or vendor statementGoogle says on-demand loading reduces input tokens and latency and helps maintain tool-selection accuracy as the tool library grows.[1]
- Primary-source or vendor statementThe feature works with Gemini 3.6 and later models and with tools exposed by remote MCP servers.[1]
- Primary-source or vendor statementClient-side retrieval is supported by setting `"execution": "client"` on the search tool, and adds one round trip per search.[1]
- Primary-source or vendor statementTool Retrieval has no separate charge; input and output are billed at standard token rates.[1]
- Primary-source or vendor statementThe announcement reports no quantitative figures for token savings, latency reduction or tool-selection accuracy.[1]
Sources
- 1Introducing Tool Retrieval in the Gemini API●blog.googleretrieved Nov 19, 2026
Updates & corrections
- First published · published automatically after passing every check
Found an error? Report it.