Google Adds Tool Retrieval to Gemini API
Google on Nov. 19 released Tool Retrieval in preview for the Gemini Interactions API, letting agents defer tool schemas and load them only when needed. There is no separate fee, and it works with Gemini 3.6 and later models, including all Gemini 4 models.
Google on Thursday, Nov. 19, added Tool Retrieval to the Gemini API, a preview feature that lets AI agents keep large tool libraries on standby and load a tool's full definition only when a task calls for it. The company announced the feature in a blog post by Lucia Loher, a product manager on the Gemini API, and Nora Kassner, a research scientist at Google DeepMind.

Key takeaways
- Developers add a tool_search tool and mark other tools with defer_loading: true; the model sees only their names and short descriptions until it needs one.
- There is no separate charge, but each server-side search triggers an extra model call billed at the input-token rate.
- The preview runs only in the Interactions API (v1beta) and works with Gemini 3.6 and later, including all Gemini 4 models.
How Gemini API Tool Retrieval works
Agents that connect to dozens or hundreds of tools normally send every tool's full parameter schema with each request. Google's developer documentation says "a few dozen function declarations can consume tens of thousands of tokens on every request," and that tool-selection accuracy drops as the list grows.
With Tool Retrieval, deferred tools stay out of context except for a name and description. When the model needs one, it asks for it by name, the API loads the schema and the model issues the function call, all inside a single request. The response lists those steps as tool_search_call, tool_search_result and function_call, which Google says makes debugging easier.
Google runs the search on its servers, so developers do not build or host an index. Setting defer_loading on a remote MCP server defers every tool that server exposes.
Client-side search for custom tool registries
Teams with their own registry, embedding index or per-user permission filter can instead declare the search tool with "execution": "client". Their app runs the lookup and returns matching schemas, which adds one round trip per search. Tools in this mode do not need to be declared upfront, so new ones can be exposed mid-conversation.
The two modes cannot be combined in one request, and only one tool_search entry is allowed per request. Loaded tools stay in context for the rest of the conversation and cannot be unloaded.
Gemini Tool Retrieval pricing and limits
Tokens are billed at the standard rates of the model being called. Deferred tools add only their name and short description to input tokens, and parameter schemas are not billed until loaded. Each server-side search, however, runs an extra model call over the full context, reported as total_tool_use_tokens.
Google is candid about the trade-off. Its documentation says static loading suits "fewer than 20 stable tools," and that with small libraries or very long contexts that search on every turn, static loading "can cost less." Google did not publish figures for token savings or accuracy gains, telling developers to measure against their own traffic.
The feature is listed for gemini-4-carbon, gemini-4-neon, gemini-4-hydrogen, gemini-4-argon, gemini-3.8-flash and gemini-3.6-flash. The older generateContent method does not support it.
Context: a pattern Anthropic shipped a year earlier
The design closely mirrors the Tool Search Tool Anthropic introduced for Claude on Nov. 24, 2025, which also uses a defer_loading: true flag. In its announcement, Anthropic said a five-server MCP setup with 58 tools consumed about 55,000 tokens before a conversation began, and reported an 85% reduction in token use in its example. Internal MCP evaluations improved from 49% to 74% for Claude Opus 4 and from 79.5% to 88.1% for Opus 4.5, the company said.
Google's release follows a busy month for the Gemini 4 family, whose models are aimed largely at long-running agent work.
What's next
The feature is in preview, and Google's documentation warns the API surface may change before general availability. Independent measurements of cost and accuracy on real agent workloads will show whether the savings hold beyond large tool libraries.
FAQ
Does Tool Retrieval cost extra? No separate fee. Standard token rates apply, including tokens used by each server-side search call.
Can I use it with generateContent? No. It is available only in the Interactions API under the v1beta version.
When should I skip it? Google suggests static loading for fewer than 20 stable tools, or when long contexts would trigger a search on every turn.
References
- Introducing Tool Retrieval in the Gemini API — Google (Nov. 19, 2026)
- Tool retrieval — Gemini API documentation, Google (accessed Nov. 21, 2026)
- Introducing advanced tool use for Claude — Anthropic (Nov. 24, 2025)
Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.
Comments 0
Comments are currently disabled.