Google Gen AI¶

GoogleGenAIModel uses the official google-genai SDK and asynchronous content streaming.

Install¶

python -m pip install "agentmuru[google]>=0.3,<0.4"
$env:GOOGLE_API_KEY = "your-key"

The client is created on the first model turn, so importing an application does not require a credential or request the network.

Configure¶

from agentmuru import Agent, Application
from agentmuru.integrations.google import GoogleGenAIModel

agent = Agent(
    name="google-assistant",
    instructions="Use tools for external facts.",
    model=GoogleGenAIModel(model="gemini-3.5-flash"),
    model_settings={
        "temperature": 0.2,
        "max_output_tokens": 800,
    },
)

application = Application(agent=agent, title="Google Agent")

Run the example:

muru run examples.providers.google_agent:application

Request translation¶

System messages and application instructions become system_instruction. User messages use the user role, while assistant messages use model. Prior tool calls become function_call parts and tool results become function_response parts.

AgentMuru sends function declarations through the generation config and disables SDK automatic function calling. This keeps permission, approval, execution, and persistence inside the AgentMuru runtime.

Function-call parts map to ToolCall after object validation. When Google omits a call ID, the adapter creates a stable local ID for that stream. Final usage metadata supplies prompt and candidate token counts.

Provider options¶

Google-specific provider_options merge into the generation config. Normalized stop maps to stop_sequences. Normalized tool choice maps to the function-calling config.

Pass api_key directly or inject an initialized client when application configuration should not use GOOGLE_API_KEY.

See the Gemini models documentation for current model availability, regions, and quotas.