The Gemini API is Google’s hosted interface to its Gemini models. Its documentation on tools describes them as specific capabilities — such as Google Search or code execution — that a model can use to answer queries, and presents them as what lets Gemini models take action in the world, access real-time information and perform complex computational tasks. Tools can be used both in standard request-response interactions and in real-time streaming sessions through the API’s Live API.
The API offers two kinds of tool: a suite of fully managed, built-in tools, and custom tools that the developer defines using function calling. For multi-step, goal-oriented systems the documentation points to a separate agents overview rather than to the tools themselves.
Capabilities
The built-in tools are listed as follows, each with the use it is intended for:
- Google Search — grounding responses in current events and facts from the web, with the stated aim of reducing hallucinations.
- Google Maps — building location-aware assistants that find places, give directions and provide local context.
- Code Execution — letting the model write and run Python code to solve maths problems or process data accurately.
- URL Context — directing the model to read and analyse content from specific web pages or documents.
- Computer Use (preview) — letting Gemini view a screen and generate actions that interact with web browser interfaces, executed on the client side (see Computer Use).
- File Search — indexing and searching the developer’s own documents to enable retrieval-augmented generation.
How a tool call is executed depends on which kind it is — the split described under
Client Tools and Server Tools. For the built-in tools other than Computer Use, the whole
process happens within one API call: Gemini decides it needs a tool, executes it on Google’s
servers — the documentation’s example searches for a stock price and then runs Python to take its
square root — and returns a final answer grounded in the results. For custom tools and for Computer
Use, the application handles execution: Gemini may return structured JSON calling a specific
function, always carrying a unique id; the application runs the function and sends the result back
with the same id; and Gemini uses it to produce a final response or another tool call. A separate
preview endpoint, gemini-3.1-pro-preview-customtools, is offered for developers building with a
mix of bash and custom tools.
As a preview feature, Gemini 3 series models can combine built-in and custom tools in a single turn, which the documentation calls tool context circulation. The developer enables a combination flag alongside the declared tools; Gemini executes built-in tools and yields to the application when it generates a client-side function call, returning the tool-call confirmation, the built-in tool’s results, the JSON for the custom call and encrypted thought signatures that preserve context. The application then returns all parts of that response together with its own function results, and Gemini generates the final answer from the combined context.
The documentation separates function calling from structured outputs by purpose: function calling is for when the model needs to perform an intermediate step by connecting to the developer’s own tools or data systems, while structured outputs are for when the model’s final response must strictly follow a specific schema, for example to render a custom interface. For Gemini 3 series models, as a preview feature, structured outputs can be combined with built-in tools so that responses grounded in external data or computation still adhere to a strict schema.
Changes introduced for Gemini 3
When Gemini 3 became available to developers, Google described in New Gemini API updates for Gemini 3 a set of API updates intended to give developers more control over how the model reasons, how it processes media and how it interacts with the outside world:
thinking_level— a parameter, from Gemini 3 onwards, that controls the maximum depth of the model’s thinking before it responds. The model treats the levels as relative guidelines for reasoning rather than strict token guarantees; “high” is suggested for complex tasks and “low” for latency- and cost-sensitive applications (see Reasoning Effort).media_resolution— configures how many tokens are used for image, video and document inputs, set per media part or globally, with defaults based on the media type when it is left unspecified. Higher resolutions improve the model’s ability to read fine text or identify small details, at the cost of more tokens and latency.- Enforced Thought signatures — starting with Gemini 3, the API enforces the return of these encrypted representations of the model’s thought process. Function calling validates them strictly on the current turn, and image generation and editing for all model parts, with a missing signature producing a 400 error; text and chat generation does not strictly enforce them, though omitting them degrades reasoning and answer quality. The official SDKs handle them automatically when standard chat history is used.
- Hosted tools with structured outputs — Grounding with Google Search and URL context can be combined with structured outputs, which the post presents as useful for agents that fetch live information from the web or specific pages and extract it into a precise JSON format.
Adoption & Ecosystem
The tools documentation links onward to guides for each built-in tool, to a function calling guide and to a tool combination guide, and notes that some tools carry costs of their own set out on the API’s pricing page.