Google has launched Tool Retrieval in preview for the Gemini API. With it, AI agents built on Gemini can keep a large library of tools on standby and pull in only the ones a task actually needs. The company says this cuts input token costs and latency, and keeps the model choosing the right tool as an agent's toolkit grows.
Lucia Loher, a product manager on the Gemini API, and Nora Kassner, a research scientist at Google DeepMind, announced the feature on November 19. It is available only in the Interactions API and works with Gemini 3.6 and later models.
Most people won't see this directly. It matters because it affects how well, and how cheaply, the agents inside Android apps and services can do multi-step work for you. An assistant that books, refunds, searches and files things across many services relies on exactly this kind of tool calling.
The problem Google is solving

Agents work by calling “tools”: functions a developer exposes, such as looking up an order or issuing a refund. Each tool comes with a parameter schema that describes its inputs. Until now, a developer generally had to send every schema with every request.
Google says this gets expensive once an agent is connected to dozens or hundreds of tools. The extra schemas raise input token counts and add latency. They also make the model worse at picking the right tool, because it has more candidates to choose from.
How Tool Retrieval works

Developers add a new tool_search tool to a request and set defer_loading: true on any tools they want to hold back. The model still sees the name and description of every deferred tool, so it knows what's available. The full parameter schemas stay out of its context until it asks for one.
When the model needs a tool, it asks for it by name. The API then loads that tool's full schema, and the model calls the function, all within the same request. Google's example is a refund agent: in one API call, the model searches for an issue_refund tool, the schema loads, and the model makes the function call. All three steps are logged in the interaction record, which should help developers find out why a model picked, or skipped, a particular tool.
Other details from Google's post:
- No index to build or host: Gemini runs the search on Google's servers. The only code changes are the new tool and the deferral flag.
- Mixed loading: Developers can keep their most-used tools loaded up front and defer only the rarely used ones.
- Prefix caching: Deferred schemas are left out of the start of the prompt, and loaded schemas are added later in the conversation, so the cached prefix doesn't change between turns.
- MCP servers: Setting the flag on a remote MCP server tool defers every tool that server exposes, so developers don't have to list each one.
- Client-side option: Teams with their own tool registry, embedding index or per-user permission filters can run the search in their own app. This costs one extra round trip per search, but it allows new tools to be added mid-conversation, for example when a user connects a new integration.
Pricing and supported models

Tool Retrieval has no separate fee. Tokens are billed at the standard rate for whichever model is called. Turning on tool_search adds a small, fixed block of input tokens. Each schema the model loads is then billed as input for the rest of the conversation, just like a tool declared up front.
Google is clear that the savings depend on the workload. Developers save tokens when the schemas they never load outweigh the ones they do, which the company says is usually the case with large libraries. It recommends measuring the effect against real traffic.
The feature works with Gemini 3.6 and later models, including the whole Gemini 4 family. It is not available in the older generateContent API.
How it fits Google's agent push

The Interactions API is Google's newer, stateful interface for building agents. It launched as a public beta in December 2025. Google moved it to general availability this June and made it the primary interface for Gemini models and agents, and new agent features have been arriving there first ever since. Tool Retrieval follows that pattern.
Google isn't the first to take this approach. Anthropic added a Tool Search Tool with the same defer_loading flag to its Claude API as a beta in November 2025. Google's version brings Gemini in line with that design and adds server-side MCP deferral and a client-side search mode.
What happens next

As of November 20, 2026, Tool Retrieval is in preview. Google hasn't said when it will become generally available. Developers can try it now in the Interactions API through Google's documentation and give feedback on the Gemini developer Discord.
Android users won't notice the change right away. Any benefit will show up gradually, as developers building Gemini-powered agents into apps and services move to the Interactions API and connect those agents to more tools.