Migration Guide: Google ADK to Local LLMs

1. Virtual Environment Dependencies

To ensure the agent runs successfully in an isolated Python virtual environment (.venv), the following dependencies must be installed. This prevents relying on global system packages and guarantees that the ADK server processes have access to the correct libraries.

Terminal Commands:

# Ensure your virtual environment is active
source .venv/bin/activate

# Install required packages
pip install litellm
pip install numpy
pip install orjson

Package Breakdown:

litellm: The core wrapper library that translates standard LLM requests into formats compatible with 100+ different providers.
numpy: Required specifically for your roll_die() tool function to generate random integers.
orjson: A high-performance JSON parsing library required under the hood by LiteLLM and the OpenAI API client to handle connection schemas.

2. Agent Configuration Code

The original script was hardcoded to use model='gemini-2.5-flash'. To route requests to a local proxy, we updated the Python code to utilize the LiteLlm model wrapper class.

Key Modifications:

Imported LiteLlm: Added from google.adk.models.lite_llm import LiteLlm to the top of the file.
Nested Configuration: Shifted the api_base and api_key variables inside the LiteLlm(...) instantiation block so they are correctly parsed as model configurations, rather than invalid agent arguments.
Provider Prefix: Appended openai/ to the model name (resulting in openai/production-deep-context) to explicitly instruct LiteLLM to use the OpenAI API translation schema when talking to your proxy on port 4000.

Final `agent.py` Code:

from google.adk.agents.llm_agent import Agent
from google.adk.models.lite_llm import LiteLlm
import numpy as np

def roll_die() -> int:
    """Return the value of a rolled dice"""
    value = np.random.randint(1, 7)
    return {"status": "success", "value": value}

root_agent = Agent(
    model=LiteLlm(
        model="openai/production-deep-context",
        api_base="http://10.11.11.200:4000/v1",
        api_key="sk-8n9_TCUmCFayrUIXK5z_ZA"
    ),
    name="dice_agent",
    description=(
        "hello world agent that can roll a dice of 6 sides"
    ),
    instruction="""
      You roll dice and tell the outcome of the dice.
    """,
    tools=[
        roll_die,
    ],
)

Migrating Google ADK to Local LLMs

1. Virtual Environment Dependencies

Terminal Commands:

2. Agent Configuration Code

Key Modifications:

Final agent.py Code:

Final `agent.py` Code: