adk-python -client

This MCP client demonstrates how the ADK (Agent Development Kit) acts as an MCP client to a flight search MCP server. It uses dynamic tool discovery to learn which capabilities the MCP server exposes (such as flight search, price checks, or multi-city routing) and then invokes those tools in a type-safe way through MCP function calls. The integration relies on ADK’s LLM-driven agent orchestration to select the appropriate tool, call it with structured inputs, and incorporate the results into the ongoing conversation with the user. Expect an asynchronous, event-driven flow where the agent maintains session context and efficiently shares data between the Gemini-based LLM and the external flight data services offered by the MCP server.

Prerequisites:

• Python 3.8+ installed
• Access to Google Gemini API (or equivalent GenAI service) with API key
• SerpAPI key for live flight data

Step 1: Create and activate a virtual environment

python -m venv venv
# macOS/Linux
source venv/bin/activate
# Windows
venv\Scripts\activate

Step 2: Install required packages

pip install google-adk            # ADK framework
pip install mcp-flight-search      # MCP server client package
pip install google-generativeai python-dotenv  # GenAI SDK and dotenv support

Step 3: Set environment variables

export GOOGLE_API_KEY="your-google-api-key"   # Gemini API key
export SERP_API_KEY="your-serpapi-key"         # SerpAPI key

Step 4: Run the MCP client (this repository)

# From project root
python -m adk_python_mcp_client

Tips and considerations:

• Ensure your Google Gemini (GenAI) access is properly provisioned; the client uses the Gemini LLM through the Google Generative AI SDK.
• The MCP server (e.g., mcp-flight-search) must be reachable from the client’s environment and must expose flight-related tools compatible with ADK tool invocation.
• Use the dynamic tool discovery feature to gracefully handle breaking changes on the MCP server—if new tools appear or existing ones change, the client should adapt without code changes.
• Manage session state with InMemorySessionService or swap in a persistent store for longer conversations.
• Properly handle resource lifecycle: ensure clean shutdowns, exit stacks, and closing of MCP connections to avoid leaks.
• If you encounter tool invocation errors, enable verbose logging in ADK to inspect tool schemas, input shapes, and return types from the MCP server.