On November 19 Google launched Tool Retrieval in preview in the Gemini Interactions API. Developers add a tool_search tool and set defer_loading: true on tools they want to hold back; the model sees only tool names and descriptions, then loads a full parameter schema by name and calls it within the same request, cutting input tokens and latency and keeping tool selection accurate as libraries grow. Remote MCP servers can be deferred as a whole, there is no separate charge, and it works with Gemini 3.6 and later models, including all Gemini 4 models.
Key Takeaways
- ✓Official: Google blog post on November 19 by Lucia Loher, Product Manager for the Gemini API, and Nora Kassner, Research Scientist at Google DeepMind, also posted by @GoogleAIStudio
- ✓Usage: add {"type": "tool_search"} to the request and set defer_loading: true on tools to hold back; frequently used tools can stay loaded up front
- ✓Flow: search, schema load and the function_call all happen inside one API call and are recorded in interaction.steps for debugging
- ✓MCP and client mode: set defer_loading on an mcp_server tool to load all of its tools on demand, or use "execution": "client" to plug in your own tool registry or permission filter at the cost of one extra round trip per search
- ✓Pricing and models: no separate charge; enabling it adds a small fixed block of input tokens and each loaded schema is billed as input for the rest of the conversation; works with Gemini 3.6 and later models, in the Interactions API only

Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
When an agent connects to dozens or hundreds of tools, sending every full parameter schema on every request inflates input tokens and latency, and tool selection gets less accurate as the candidate list grows. Google's Tool Retrieval lets an agent keep only the tool list in context and load full schemas when a task actually needs them.
How it works
You add a tool_search tool and set defer_loading: true on the tools to hold back. The model still sees each deferred tool's name and description, asks for one by name when needed, the API loads the schema, and the model emits the function_call, all in one request and recorded in interaction.steps. Search runs on the server with no index to build or host, and loaded schemas are appended later in the conversation so the cached prefix is preserved. Remote MCP servers can be deferred as a whole, and teams with their own tool registry can use client-side mode with "execution": "client", which adds one round trip per search.
Benchmarks
Google published no benchmark scores or savings figures, only the billing model: there is no separate charge, 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 more schemas go unloaded, the more you save, and Google advises measuring against your own traffic.
What developers should do
The feature is in preview, available only in the Interactions API, for Gemini 3.6 and later models including all Gemini 4 models. Teams with large tool libraries can keep high-frequency tools loaded, defer the long tail, and compare token usage and tool-selection accuracy before and after; projects connected to a large MCP server can simply defer the whole server.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.