Google has introduced a new capability called Tool Retrieval in its Gemini API, currently available in preview through the Interactions API. This feature addresses the challenge of managing large libraries of software tools by deferring the loading of detailed parameter schemas until they are explicitly needed for a task, optimizing performance and resource usage in AI agent workflows.
In conventional setups, connecting an AI agent to dozens or hundreds of tools requires transmitting every tool’s full parameter schema with each request. This leads to increased input token consumption, slower response times, and diminished tool selection accuracy as the list of candidates grows. Tool Retrieval resolves these issues by letting the agent maintain a standby tool library that only loads schemas relevant to the current request.
To use Tool Retrieval, developers add a tool_search tool to their requests and mark tools with defer_loading set to true. The agent retains access to basic information such as each tool’s name and description, keeping them visible without their parameter details. When the agent decides a specific tool is required, it requests the schema by name mid-request, enabling dynamic loading and immediate function invocation without additional round-trips.
This feature also integrates seamlessly with caching strategies by keeping deferred schemas out of the initial prompt to preserve cached prefix efficiency. Furthermore, Tool Retrieval supports remote MCP servers, allowing entire servers’ toolsets to be loaded on demand without manually listing individual tools, simplifying integration of large external tool repositories.
For clients maintaining their own tool registries or permission filters, Tool Retrieval supports client-side search execution, enabling dynamic tool discovery and permitting the exposure of new tools during a conversation without upfront declaration. This extends flexibility for personalized or evolving AI agent capabilities.
Importantly, Tool Retrieval does not incur additional charges beyond the standard input and output token billing for Gemini models. Latency reductions and token savings arise especially when a large portion of deferred schemas remain unused during sessions. Currently, Tool Retrieval is compatible with Gemini 3.6 and subsequent versions, including all Gemini 4 models, accessible via the Interactions API.
By implementing tool parameter loading on demand, Google’s Tool Retrieval feature is set to enhance the scalability, responsiveness, and efficiency of AI-powered agents, particularly those interacting with extensive tool ecosystems. Developers interested in adopting this capability can access detailed documentation, model comparisons, and community discussions through Google's official channels.
While the feature offers clear benefits in token cost savings and latency reduction, its efficacy depends on usage patterns. Large libraries with frequent accesses to most tools may see less advantage. Additionally, client-side retrieval introduces an extra network round-trip per tool search, which should be accounted for in latency-sensitive applications.
Source & further reading
Introducing Tool Retrieval in the Gemini API ↗Source published 2026-11-19. Claims and limitations should be checked against the original announcement or research.