Gemini API Quickstart: From API Key to Your First Agent Call

Get an API Key

Open Google AI Studio and create a key at aistudio.google.com/apikey. Store it in an environment variable rather than in source code:

export GEMINI_API_KEY="your-key-here"

On Windows PowerShell:

$env:GEMINI_API_KEY="your-key-here"

Install the SDK

Python:

pip install -U google-genai

Node.js or TypeScript:

npm install @google/genai

The current SDKs replace the older google-generativeai and @google/generative-ai packages. Use the new ones for new projects.

Your First Call

Python:

from google import genai

client = genai.Client()
response = client.models.generate_content(
    model="gemini-flash-latest",
    contents="Explain the Model Context Protocol in two sentences.",
)
print(response.text)

TypeScript:

import { GoogleGenAI } from "@google/genai";

const ai = new GoogleGenAI({});
const response = await ai.models.generateContent({
  model: "gemini-flash-latest",
  contents: "Explain the Model Context Protocol in two sentences.",
});
console.log(response.text);

The client reads GEMINI_API_KEY automatically. If you need to pass it explicitly, use the apiKey option.

Streaming Responses

Streaming keeps interfaces responsive when outputs are long:

for chunk in client.models.generate_content_stream(
    model="gemini-flash-latest",
    contents="Write a migration plan for moving a monolith to services.",
):
    print(chunk.text, end="")

Function Calling

Function calling is how Gemini becomes an agent. Declare tools, let the model choose one, execute it in your code, and return the result:

from google.genai import types

def get_weather(city: str) -> str:
    return f"18C and clear in {city}"

tools = [get_weather]
response = client.models.generate_content(
    model="gemini-flash-latest",
    contents="What is the weather in Bengaluru?",
    config=types.GenerateContentConfig(tools=tools),
)
print(response.text)

Keep tool functions small, validate every argument, and never pass raw model output into a shell or SQL query.

Choosing a Model

Use the Flash models for high-volume extraction, classification, and chat. Use the Pro models for complex reasoning, long documents, and planning. Model aliases such as gemini-flash-latest move forward automatically; pin an explicit version in production if you need reproducible behavior.

Cost and Limits

The Gemini API has a free tier with rate limits and a paid tier billed per token. Free-tier quotas change frequently, so check the current pricing page before sizing a workload. For production, add retries with exponential backoff, cap output tokens, and cache repeated prompts where possible.

Production Checklist

• Keep the key server-side; never ship it in a browser bundle
• Set per-request timeouts and retry on 429 and 5xx responses
• Log token usage and latency per route
• Validate structured output against a schema before using it
• Use a pinned model version for reproducible evaluations

Next Steps

Once the first call works, wrap it in a tool-using loop and connect real data. The MCP guide on Tecrese shows how to expose those tools to any compatible agent, and the AI Launchpad has copy-paste setup commands for the whole stack.