Google introduced Tool Retrieval in preview in the Gemini Interactions API on 19 November 2026, letting developers hold back the detailed instructions for an agent’s tools until the agent needs them. Google says the feature is designed for agents connected to dozens or hundreds of tools, where sending every tool’s full specification with each request increases input tokens and latency.
Key points
- Developers add tool_search and mark selected tools with defer_loading: true.
- The model sees deferred tools’ names and descriptions, then loads their parameter schemas when needed.
- Google says the feature supports Gemini 3.6 and later models, including all Gemini 4 models.
- There is no separate Tool Retrieval charge, though loaded schemas count towards input tokens.
Tool search keeps schemas on standby
A parameter schema is a tool’s instruction sheet: it sets out the inputs a function expects, such as an order ID and an amount for the refund function in Google’s example. With Tool Retrieval, the model initially sees each deferred tool’s name and description, but the schema stays outside its working context. When the model requests a tool by name, Google says the API supplies its schema and the model makes the function call within the same request.
Developers enable that sequence by adding a tool_search tool to the request and setting defer_loading: true on the tools they want to hold back. Google says the search, schema retrieval and function call appear in interaction.steps, where developers can inspect them when debugging tool names or descriptions. The application still executes the requested function and returns its result.
The Interactions API reached general availability in mid-2026 and keeps state on Google’s servers across multi-step tasks, Crypto Briefing reported.
Gemini 3.6 support and token costs
Google says Tool Retrieval works in the Interactions API with Gemini 3.6 and later models, including all Gemini 4 models. Its example uses Gemini 4 Hydrogen, whose launch alongside Antigravity model switching AI Affairs previously reported.
Google says Tool Retrieval carries no separate charge: input and output tokens are billed at the called model’s standard rates. Enabling tool_search adds a small fixed amount of input text, and a loaded schema counts as input for the remainder of the conversation. Google says token use falls when the schemas left unloaded outweigh those retrieved, and advises developers to measure the result against their own traffic.
MCP servers and client-side searches
For a remote MCP server, Google says setting defer_loading: true on the mcp_server tool makes every tool exposed by that server available for loading on demand. Google also says deferred schemas arrive later in a conversation, leaving the beginning of a prompt unchanged for prefix caching.
Applications with their own tool registry or permission filter can instead set tool_search execution to client and perform the lookup themselves, according to Google. That option lets an application introduce tools partway through a conversation, but adds one round trip for each search because the application runs it.