Generative AI Google Brain

agents-agent-engine-tutorial-claude-with-adk-on-agent-engine

3 min read

Claude with ADK on Vertex AI Agent Engine

Source notebook

Repo path: agents/agent_engine/tutorial_claude_with_adk_on_agent_engine.ipynb · Open on GitHub · advanced

Builds and deploys a Claude-powered ADK Reddit agent on Vertex AI Agent Engine with memory and tracing.

Summary

This notebook teaches how to build a Google ADK agent using Anthropic Claude on Vertex AI, first as a local social media assistant and then with Reddit tools through MCP. It demonstrates local testing with an ADK Runner, deployment to Vertex AI Agent Engine as an AdkApp, session-based remote chat, generated memories, and an optional update that adds Memory Bank retrieval and tracing.

Key code patterns

Register Claude for ADK

from google.adk.models.anthropic_llm import Claude
from google.adk.models.registry import LLMRegistry
 
MODEL_ID = "claude-sonnet-4@20250514"
LLMRegistry.register(Claude)

Registers the Anthropic Claude model class so ADK agents can use the Vertex-hosted Claude model id.

Local ADK agent

root_agent = LlmAgent(
    name="SocialMediaAssistant",
    model=os.getenv("MODEL_ID", "claude-sonnet-4@20250514"),
    instruction="You are a creative and knowledgeable Social Media Assistant.",
    tools=[],
)

Defines the agent identity, system instruction, model, and tool list before local or remote execution.

MCP Reddit toolset

MCPToolset(
    connection_params=StdioConnectionParams(
        server_params=StdioServerParameters(command="mcp-reddit"),
    ),
    errlog=errlog,
)

Connects the ADK agent to an external Reddit MCP server and redirects errors for Colab compatibility.

Deploy ModuleAgent

remote_app = agent_engines.create(
    display_name="reddit_assistant_agent",
    agent_engine=agent_engines.ModuleAgent(
        module_name="root_agent",
        agent_name="agent_app",
    ),
    extra_packages=["root_agent.py", "installation_scripts/install_local_mcp.sh"],
)

Packages the ADK app module and installer script for managed deployment on Vertex AI Agent Engine.

Memory and tracing

agent_app = AdkApp(
    agent=root_agent,
    session_service_builder=session_service_builder,
    memory_service_builder=memory_service_builder,
    enable_tracing=True,
)

Adds Agent Engine tracing and connects the deployed app to Vertex AI Memory Bank.

Models & APIs used

  • Models: claude-sonnet-4@20250514
  • APIs / services: Vertex AI, Vertex AI Agent Engine, Vertex AI Memory Bank, Cloud Storage
  • SDKs / libraries: google-cloud-aiplatform, google-adk, anthropic, vertexai, google-genai, litellm, fastmcp, redditwarp, praw

When to use this

Use this pattern to deploy an ADK agent powered by Claude on Vertex AI Agent Engine with MCP tools, sessions, memory, and tracing.

Gotchas & caveats

  • The notebook requires an existing Google Cloud project with the Vertex AI API enabled.
  • A Cloud Storage bucket is created and used as the Vertex AI staging bucket for deployment artifacts.
  • The location is set to europe-west1 throughout the notebook.
  • Reddit client id, client secret, and refresh token are required for the Reddit MCP tool.
  • The mcp-reddit server is installed from GitHub with uv through a custom shell installation script.
  • The notebook redirects MCP tool stderr to an aiofiles errlog in Colab to avoid a fileno error.
  • Environment variables are passed into the deployed Agent Engine app for model and Reddit credentials.
  • The cleanup section explicitly deletes the Agent Engine deployment and Cloud Storage bucket to avoid charges.

Best practices

  • Test the ADK agent locally with InMemorySessionService and Runner before deploying.
  • Package the deployable agent in a root_agent.py module and expose an AdkApp entry point.
  • Use a VertexAiSessionService builder for production sessions on Agent Engine.
  • Pass runtime credentials and model settings through environment variables for the deployed app.
  • Use build_options installation steps to install external MCP runtime dependencies during deployment.
  • Use generate_memories or async_add_session_to_memory to persist useful facts from prior sessions.
  • Use PreloadMemoryTool to retrieve relevant Memory Bank facts at the start of each turn.
  • Delete deployed Agent Engine resources and staging buckets when finished.

Graph View

Backlinks