Google on November 19, 2026 launched Tool Retrieval in the Gemini API, available in preview in the Interactions API. The blog post was written by Lucia Loher, a Product Manager for the Gemini API, and Nora Kassner, a Research Scientist at Google DeepMind. It targets a practical scaling problem for agents: when an agent is connected to dozens or hundreds of tools, sending every full parameter schema with every request inflates input tokens, adds latency and lowers tool-selection accuracy as the candidate list grows.
With Tool Retrieval, an agent keeps its full tool library on standby and loads a parameter schema only when the current task needs it. A developer adds a tool of type tool_search to the request and sets defer_loading to true on the tools to hold back. The model still sees each deferred tool's name and description, so it knows what exists, but the schemas stay out of context. When it needs a tool, it asks for it by name, the API loads the schema, and the model emits the function call, all inside one request. The three steps, tool_search_call, tool_search_result and function_call, appear in the interaction steps so developers can see which tools the model considered.
Google says there is no index to build or host because Gemini runs the search on the server. Developers can keep frequently used tools loaded up front and defer the long tail. The company also says the design works with prefix caching, since deferred schemas stay out of the start of the prompt and loaded schemas are added later in the conversation, so the cached prefix remains the same between turns.
Remote MCP servers are supported too. Setting defer_loading on an mcp_server tool makes every tool that server exposes load on demand, which lets teams connect a large server without listing its tools one by one. For teams with their own registry, embedding index or per-user permission filter, a client-side mode declares tool_search with execution set to client and defines custom search parameters. That mode adds one round trip per search, because the application performs the lookup, and it allows new tools to be exposed mid-conversation when a user connects an integration.
On cost, Google states that Tool Retrieval has no separate charge. Input and output tokens are billed at the standard rates for the model used. Enabling tool_search adds a small, fixed block of input tokens, and each schema the model loads is billed as input for the rest of the conversation, the same as a tool declared up front. Savings appear when the schemas never loaded outweigh those that are, which Google calls the usual case for large libraries, and it advises developers to measure against their own traffic.
The feature works with Gemini 3.6 and later models, including all Gemini 4 models, in the Interactions API. The sources reviewed give no accuracy or token-saving benchmarks, so the size of the benefit is untested outside Google.