Google puts Tool Retrieval into preview in the Gemini API
Agents can now keep large tool libraries on standby and load parameter schemas only when a task needs them.
Google has put Tool Retrieval into preview in the Gemini API, inside the Interactions API. Announced on Nov 19, 2026, it targets agents that connect to many tools. Rather than sending every parameter schema on every request, the model keeps its library on standby and loads schemas only when a task needs them. Google's announcement post presents Gemini API tool retrieval as a way to cut input tokens and latency while keeping tool selection accurate as the candidate list grows.
What the source confirms
Two changes enable it: add a tool_search tool to the request, and set defer_loading: true on the tools you want held back. Deferred tools stay visible by name and description, so the model knows what exists, but their parameter schemas are kept out of context. When the model needs one it asks for it by name, the API loads the full schema, and the model emits the function call, all inside the same request.
The three steps, a tool_search_call, a tool_search_result and the function_call, show up in interaction.steps. Google pitches that as a debugging aid for cases where tool names or descriptions cause bad selections.
There is no index to build or host; the search runs on Google's servers, so the code changes are the new tool entry and the flag. Google's advice is to leave frequent tools undeferred and defer the long tail.
Caching, MCP servers and client-side search
Deferred schemas stay out of the start of the prompt and loaded schemas arrive later in the conversation, so the cached prefix stays stable from turn to turn. Prefix caching keeps working.
For remote MCP servers in the Gemini API, setting defer_loading: true on an mcp_server tool loads every tool that server exposes on demand, without listing them one by one.
Teams with their own registry, embedding index or per-user permission filter can declare tool_search with "execution": "client" and define their own search parameters. The app performs the lookup, which adds one round trip per search. Tools need not be declared upfront, so new ones can appear mid-conversation when a user connects an integration.
Pricing and supported models
There is no separate charge. Tokens bill at the standard rates for the model you call. tool_search adds a small, fixed block of input tokens, and each loaded schema is billed as input for the rest of the conversation, the same as a tool declared up front. You save when the schemas you never load outweigh the ones you do, which Google says to measure against your own traffic.
Support covers Gemini 3.6 and later, including all Gemini 4 models, in the Interactions API. Google's sample uses gemini-4-hydrogen with a single deferred issue_refund tool.
Limits and open questions
The post gives no token or latency numbers, no ceiling on library size, and no date for general availability. Deferred tools still contribute a name and description to the prompt, so a very large library still carries a base cost before any schema loads, which suggests the saving shrinks at the margin. Client-side retrieval adds a round trip per search, likely eating into the latency gain. Whether prefix caching behaves the same in client mode is not addressed.
What this means for building agents
Tool count has been a practical design limit. Every tool added makes the prompt larger and the choice harder, and accuracy falls as the candidate list grows. Deferral turns a big library into something you can ship instead of a threshold to stay under, which likely changes how integrations get scoped. The step-by-step trace is arguably as useful as the token saving, since tool selection failures usually trace back to vague names and descriptions.
For readers building with AI on Draftship: Draftship is an AI app builder where you describe an app in plain language and it writes the code, builds it and hosts it, with a built-in database, user sign-in and hosting. Paid plans add private projects, a choice of AI model, source code access, custom domains and GitHub push. Draftship is not affiliated with Google, and nothing in the announcement says Draftship supports Tool Retrieval. If you are wiring AI into your own app, the docs on adding AI to your app are the place to start.
Primary sources
- Introducing Tool Retrieval in the Gemini API · Google AI Blog, Nov 19, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.