google-adk-models

name	google-adk-models
description	Use non-Gemini models with ADK agents via LiteLLM or Anthropic. Use when wiring Claude, GPT-4, Llama, Mistral, or any LiteLLM-supported provider into an ADK agent.

Google ADK — Models (Non-Gemini)

Overview

ADK's model field accepts either a string (Gemini models) or a BaseLlm instance (any other provider via LiteLLM).

Provider	Model String Pattern	Install
Gemini (native)	"gemini-2.5-flash"	pip install google-adk
Any via LiteLLM	LiteLlm(model="provider/model")	pip install google-adk[extensions]

Install LiteLLM Support

pip install google-adk[extensions]

Using LiteLLM (Any Provider)

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

agent = Agent(
    name="claude_agent",
    model=LiteLlm(model="anthropic/claude-sonnet-4-20250514"),
    instruction="You are a helpful assistant.",
    tools=[my_tool],
)

Provider-Specific Examples

Anthropic (Claude)

import os
os.environ["ANTHROPIC_API_KEY"] = "sk-..."

agent = Agent(
    name="claude_agent",
    model=LiteLlm(model="anthropic/claude-sonnet-4-20250514"),
    instruction="...",
)

OpenAI (GPT)

import os
os.environ["OPENAI_API_KEY"] = "sk-..."

agent = Agent(
    name="gpt_agent",
    model=LiteLlm(model="openai/gpt-4o"),
    instruction="...",
)

Azure OpenAI

import os
os.environ["AZURE_API_KEY"] = "..."
os.environ["AZURE_API_BASE"] = "https://your-resource.openai.azure.com/"
os.environ["AZURE_API_VERSION"] = "2024-02-01"

agent = Agent(
    name="azure_agent",
    model=LiteLlm(model="azure/gpt-4o-deployment-name"),
    instruction="...",
)

Vertex AI (Claude via Google Cloud)

import os
os.environ["VERTEXAI_PROJECT"] = "your-gcp-project-id"
os.environ["VERTEXAI_LOCATION"] = "us-east5"

agent = Agent(
    name="vertex_claude",
    model=LiteLlm(model="vertex_ai/claude-3-7-sonnet@20250219"),
    instruction="...",
)

Ollama (Local Models)

agent = Agent(
    name="local_agent",
    model=LiteLlm(model="ollama/llama3.1"),
    instruction="...",
)

Additional LiteLLM Arguments

Pass extra kwargs to customize the LiteLLM completion call:

agent = Agent(
    name="custom_agent",
    model=LiteLlm(
        model="anthropic/claude-sonnet-4-20250514",
        temperature=0.7,
        max_tokens=4096,
        drop_params=True,
    ),
    instruction="...",
)

Default Model for All Agents

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

LlmAgent.set_default_model(LiteLlm(model="openai/gpt-4o"))

Multi-Model Systems

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

router = Agent(
    name="router",
    model="gemini-2.5-flash",
    instruction="Route requests to the appropriate specialist.",
    sub_agents=[fast_agent, smart_agent],
)

fast_agent = Agent(
    name="fast_worker",
    model="gemini-2.5-flash",
    instruction="Handle simple tasks quickly.",
)

smart_agent = Agent(
    name="smart_worker",
    model=LiteLlm(model="anthropic/claude-sonnet-4-20250514"),
    instruction="Handle complex reasoning tasks.",
)

Gemini Models (Native — No LiteLlm Needed)

Model	Use Case
"gemini-2.5-flash"	Fast, cost-effective (default)
"gemini-2.5-pro"	Complex reasoning
"gemini-2.0-flash"	Legacy

Key Rules

• Gemini models use plain strings: model="gemini-2.5-flash"
• All other providers use LiteLlm(model="provider/model-name")
• Environment variables for auth must be set BEFORE creating the agent
• LiteLLM model strings follow the "provider/model" convention
• Don't use LiteLLM for Gemini models — ADK warns against this (use native strings)
• drop_params=True silently ignores unsupported params (useful for cross-provider code)

Related Skills

• google-adk-llm-agent — Agent configuration (where model is set)
• google-adk-deploy — Deploying with non-Gemini models (env vars in secrets)