Imagine asking your AI model, “What’s the weather in Tokyo right now?” and instead of hallucinating an answer, it calls your actual Python function, fetches live data, and responds correctly. That’s how empowering the tool call functions in the Gemma 4 from Google are. A truly exciting addition to open-weight AI: this function calling is structured, reliable, and built directly into the AI model!

Coupled with Ollama for local referencing, it allows you to develop non-cloud-dependent AI agents. The best part – these agents have access to real-world APIs and services locally, without any subscription. In this guide, we will cover the concept and implementation architecture as well as three tasks that you can experiment with immediately.

Also read: Running Claude Code for Free with Gemma 4 and Ollama

What is Tool Calling in LLMs?

Conversational language models have a limited knowledge based on when they were developed. Hence, they can offer only an approximate answer when you ask for current market prices or current weather conditions. This lack was addressed by providing an API wrapper around common models (functions). The aim – to solve these types of questions via (tool-calling) service(s).

By enabling tool-calling, the model can recognize:

• When it is necessary to retrieve outside information
• Identify the correct function based on the provided API
• Compile correctly formatted method calls (with arguments)

It then waits until the execution of that code block returns the output. It then composes an assessed answer based on the received output.

To clarify: the model never executes the method calls that have been created by the user. It only determines which methods to call and how to structure the method call argument list. The user’s code will execute the methods that they called via the API function. In this scenario, the model represents the brain of a human, while the functions being called represent the hands.

Gemma 4 Tool Calling Architecture

Before you begin writing code, it is beneficial to understand how everything works. Here is the loop that each tool in Gemma 4 will follow, as it makes tool calls:

1. Define functions in Python to perform actual tasks (i.e., retrieve weather data from an external source, query a database, convert money from one currency to another).
2. Create a JSON schema for each of the functions you have created. The schema should contain the name of the function and what its parameters are (including their types).
3. When the system sends a message to you, you send both the tool-schemas you have created and the system’s message to the Ollama API.
4. The Ollama API returns data in a tool_calls block rather than plain text.
5. You execute the function using the parameters sent to you by the Ollama API.
6. You return the result back to the Ollama API as a ‘role’:’tool’ response.
7. The Ollama API receives the result and returns the answer to you in natural language.

This two-pass pattern is the foundation for every function-calling AI agent, including the examples shown below.

Gemma 4 Tool Calling Environment Setup

To execute these tasks, you will need two components: Ollama must be installed locally on your machine, and you will need to download the Gemma 4 Edge 2B model. There are no dependencies beyond what is provided with the standard installation of Python, so you don’t need to worry about installing Pip packages at all.

1. To install Ollama with Homebrew or MacOS:

# Install Ollama (macOS/Linux) 
curl --fail -fsSL https://ollama.com/install.sh | sh 

2. To download the model (which is approximately 2.5 GB):

# Download the Gemma 4 Edge Model – E2B 
ollama pull gemma4:e2b

After downloading the model, use the Ollama list to confirm it exists in the list of models. You can now connect to the running API at the URL http://localhost:11434 and run requests against it using the helper function we will create:

import json, urllib.request, urllib.parse
def call_ollama(payload: dict) -> dict:
    data = json.dumps(payload).encode("utf-8")
    req = urllib.request.Request(
        "http://localhost:11434/api/chat",
        data=data,
        headers={"Content-Type": "application/json"},
    )
    with urllib.request.urlopen(req) as resp:
        return json.loads(resp.read().decode("utf-8"))

No third-party libraries are needed; therefore, the agent can run independently and provides complete transparency.

Also read: How to Run Gemma 4 on Your Phone: A Hands-On Guide

Hands-on Task 01: Live Weather Lookup

The first of our methods uses open-meteo that pulls live data for any location through a free weather API that does not need a key in order to pull the information down to the local area based on longitude/latitude coordinates. If you’re going to use this API, you will have to perform a series of steps :

1. Write your function in Python

def get_current_weather(city: str, unit: str = "celsius") -> str:
    geo_url = f"https://geocoding-api.open-meteo.com/v1/search?name={urllib.parse.quote(city)}&count=1"
    with urllib.request.urlopen(geo_url) as r:
        geo = json.loads(r.read())
    loc = geo["results"][0]
    lat, lon = loc["latitude"], loc["longitude"] 
    url = (f"https://api.open-meteo.com/v1/forecast"
           f"?latitude={lat}&longitude={lon}"
           f"&current=temperature_2m,wind_speed_10m"
           f"&temperature_unit={unit}")
    with urllib.request.urlopen(url) as r:
        data = json.loads(r.read())
    c = data["current"]
    return f"{city}: {c['temperature_2m']}°, wind {c['wind_speed_10m']} km/h" 

2. Define your JSON schema

This provides the information to the model so that Gemma 4 knows exactly what the function will be doing/expecting when it is called.

 weather_tool = { 

    "type": "function",
    "function": {
        "name": "get_current_weather",
        "description": "Get live temperature and wind speed for a city.",
        "parameters": {
            "type": "object",
            "properties": {
                "city": {"type": "string", "description": "City name, e.g. Mumbai"},
                "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}
            },
            "required": ["city"]
        }
    }

3. Create a query for your tool call (as well as handle and process the response back) 

messages = [{"role": "user", "content": "What's the weather in Mumbai right now?"}] response = call_ollama({"model": "gemma4:e2b", "messages": messages, "tools": [weather_tool], "stream": False}) msg = response["message"]
if "tool_calls" in msg: tc = msg["tool_calls"][0] fn = tc["function"]["name"] args = tc["function"]["arguments"] result = get_current_weather(**args) # executed locally
messages.append(msg) 
messages.append({"role": "tool", "content": result, "name": fn})
final = call_ollama({"model": "gemma4:e2b", "messages": messages, "tools": [weather_tool], "stream": False}) 
print(final["message"]["content"])

Output

Hands-on Task 02: Live Currency Converter

The classic LLM fails by hallucinating currency values and not being able to provide accurate, up-to-date currency conversion. With the help of ExchangeRate-API, the converter can get the latest foreign exchange rates and convert accurately between two currencies.

Once you complete Steps 1-3 below, you will have a fully functioning converter in Gemma 4:

1. Write your Python function

def convert_currency(amount: float, from_curr: str, to_curr: str) -> str:
    url = f"https://open.er-api.com/v6/latest/{from_curr.upper()}"
    with urllib.request.urlopen(url) as r:
        data = json.loads(r.read())
    rate = data["rates"].get(to_curr.upper())
    if not rate:
        return f"Currency {to_curr} not found."
    converted = round(amount * rate, 2)
    return f"{amount} {from_curr.upper()} = {converted} {to_curr.upper()} (rate: {rate})"

2. Define your JSON schema 

currency_tool = { 

    "type": "function",
    "function": {
        "name": "convert_currency",
        "description": "Convert an amount between two currencies at live rates.",
        "parameters": {
            "type": "object",
            "properties": {
                "amount":    {"type": "number", "description": "Amount to convert"},
                "from_curr": {"type": "string", "description": "Source currency, e.g. USD"}, 
                "to_curr":   {"type": "string", "description": "Target currency, e.g. EUR"}
            },
            "required": ["amount", "from_curr", "to_curr"]
        } 
    }
} 

3. Test your solution using a natural language query

response = call_ollama({
    "model": "gemma4:e2b",
    "messages": [{"role": "user", "content": "How much is 5000 INR in USD today?"}],
    "tools": [currency_tool],
    "stream": False
}) 

Gemma 4 will process the natural language query and format a proper API call based on amount = 5000, from = ‘INR’, to = ‘USD’. The resulting API call will then be processed by the same ‘Feedback’ method described in Task 01.

Output

Hands-on Task 03: Multi-Tool Agent (Stacked Calls)

Gemma 4 excels at this task. You can supply the model multiple tools simultaneously and submit a compound query. The model coordinates all the required calls in one go; manual chaining is unnecessary.

1. Add the timezone tool

def get_current_time(city: str) -> str: 

    url = f"https://timeapi.io/api/Time/current/zone?timeZone=Asia/{city}"
    with urllib.request.urlopen(url) as r:
        data = json.loads(r.read())
    return f"Current time in {city}: {data['time']}, {data['dayOfWeek']} {data['date']}"
time_tool = {
    "type": "function",
    "function": {
        "name": "get_current_time",
        "description": "Get the current local time in a city.",
        "parameters": {
            "type": "object",
            "properties": {
                "city": {"type": "string", "description": "City name for timezone, e.g. Tokyo"}
            },
            "required": ["city"]
        }
    } 

2. Build the multi-tool agent loop

TOOL_FUNCTIONS = { "get_current_weather": get_current_weather, "convert_currency": convert_currency, "get_current_time": get_current_time, } 

def run_agent(user_query: str): all_tools = [weather_tool, currency_tool, time_tool] messages = [{"role": "user", "content": user_query}] 

response = call_ollama({"model": "gemma4:e2b", "messages": messages, "tools": all_tools, "stream": False}) 
msg = response["message"] 
messages.append(msg) 
 
if "tool_calls" in msg: 
    for tc in msg["tool_calls"]: 
        fn     = tc["function"]["name"] 
        args   = tc["function"]["arguments"] 
        result = TOOL_FUNCTIONS[fn](**args) 
        messages.append({"role": "tool]]]", "content": result, "name": fn}) 
 
    final = call_ollama({"model": "gemma4:e2b", "messages": messages, "tools": all_tools, "stream": False}) 
    return final["message"]["content"]
return msg.get("content", "")

3. Execute a compound/multi-intent query

print(run_agent(
    "I'm flying to Tokyo tomorrow. What's the current time there, "
    "the weather, and how much is 10000 INR in JPY?"
))e

Output

Here, we described three distinct functions with three separate APIs in real-time through natural language processing using one common concept. It includes all local execution without cloud solutions from the Gemma 4 instance; none of these components utilize any remote resources or cloud.

What Makes Gemma 4 Different for Agentic AI?

Other open weight models can call tools, yet they do not perform reliably, and this is what differentiates them from Gemma 4. The model consistently provides valid JSON arguments, processes optional parameters correctly, and determines when to return knowledge and not call a tool. As you keep using it, keep in mind the following:

• Schema quality is critically important. If your description field is vague, you will have a difficult time identifying arguments for your tool. Be specific with units, formats, and examples.
• The required array is validated by Gemma 4. Gemma 4 respects the needed/optional distinction.
• Once the tool returns a result, that result becomes a context for any of the “role”: “tool” messages you send during your final pass. The richer the result from the tool, the richer the response will be.
• A common mistake is to return the tool result as “role”: “user” instead of “role”: “tool”, as the model will not attribute it correctly and will attempt to re-request the call.

Also read: Top 10 Gemma 4 Projects That Will Blow Your Mind

Conclusion

You have created a real AI agent that uses the Gemma 4 function-calling feature, and it is working entirely locally. The agent-based system uses all the components of the architecture in production. Potential next steps can include:

• adding a file system tool that will allow for reading and writing local files on demand;
• using a SQL database as a means for making natural language data queries;
• creating a memory tool that will create session summaries and write them to disk, thus providing the agent with the ability to recall past conversations

The open-weight AI agent ecosystem is evolving quickly. The ability for Gemma 4 to natively support structured function calling affords substantial autonomous functionality to you without any reliance on the cloud. Start small, create a working system, and the building blocks for your next projects will be ready for you to chain together.