agent using adk, litellm and groq api-key

1. Environment Setup & Troubleshooting

First, I signed up for the Groq console at console.groq.com/home and generated an API key at console.groq.com/keys.

I used Python 3.13 to create the virtual environment:

python3.13 -m venv .venv

The Dependencies & Downgrade Workaround

Initially, I planned to run a standard install, but I ran into a LiteLLM serialization issue. To bypass this, I had to downgrade google-adk to the 1.28.0 range and include python-dotenv:

pip3.13 install "google-adk~=1.28.0" litellm python-dotenv

Updated Project Structure (ls -R)

.
├── requirements.txt
└── time_agent/
    ├── __init__.py
    └── agent.py

Inside the .env file, I added the Groq API key:

GROQ_API_KEY="your_api_key_here"

---

2. Resolving the Groq Request Error

During implementation, I hit a Groq request error regarding reasoning/thought blocks. I added this custom callback to strip out reasoning content from the LLM request parts before it hits the model:

def strip_reasoning_content(callback_context, llm_request):
    for content in llm_request.contents:
        if content.parts:
            content.parts = [
                p for p in content.parts if not (hasattr(p, "thought") and p.thought)
            ]
    return None

---

3. The Complete Code (time_agent/agent.py)

Here is the complete source integrating the tool mapping, the custom callback, and the Groq LLM configuration.

import datetime
import zoneinfo

from google.adk.agents import LlmAgent
from google.adk.models.lite_llm import LiteLlm


def strip_reasoning_content(callback_context, llm_request):
    for content in llm_request.contents:
        if content.parts:
            content.parts = [
                p for p in content.parts if not (hasattr(p, "thought") and p.thought)
            ]
    return None


def get_current_time(timezone):
    tz = zoneinfo.ZoneInfo(timezone)
    now = datetime.datetime.now(tz)
    return f"Current time in {timezone}: {now.strftime('%I:%M %p on %A, %B %d, %Y')}"


root_agent = LlmAgent(
    model=LiteLlm(model="groq/openai/gpt-oss-20b"),
    name="groq_time_assistant",
    description="An assistant powered by Groq that can tell the current time",
    instruction=(
        "You are a helpful assistant. When the user asks about the current time "
        "in any location, use the get_current_time tool with the appropriate "
        "IANA timezone name (e.g. 'US/Eastern', 'Asia/Kolkata', 'Europe/Paris'). "
        "If the user gives a city name, map it to the correct timezone yourself."
    ),
    tools=[get_current_time],
    before_model_callback=strip_reasoning_content,
)

---

4. Execution & Testing

I ran adk web to spin up the web UI and test the code locally.

Backend Logs

The console confirmed LiteLLM completion calls successfully routing through Groq:

13:56:26 - LiteLLM:INFO: utils.py:4054 - LiteLLM completion() model= openai/gpt-oss-20b; provider = groq
2026-05-31 13:56:26,332 - INFO - utils.py:4054 - LiteLLM completion() model= openai/gpt-oss-20b; provider = groq

Web UI Behavior

When tested with the query “what’s the current time in tokyo ?”, the agent successfully mapped the city to the correct timezone, called the tool, and outputted the final time.

• Thought: Need to map Tokyo to timezone: Asia/Tokyo. Then call get_current_time.
• Tool Call: get_current_time
• Output: Sure! The current time in Tokyo (Asia/Tokyo) is 05:26 PM on Sunday, May 31, 2026.

Metrics

Checking the Groq Console dashboard after the test runs showed a clean and efficient footprint with total token usage sitting at 2.1K tokens.