May 17
Loading video player...
In this session of The Average Engineer, we explore how to bypass manual tool creation by leveraging Model Context Protocol (MCP) servers. We take our existing Trip Planner agent, powered by a local Qwen LLM, and upgrade its architecture. Instead of hardcoding a "post to Atlassian" function to execute a Create Page API , we deploy a pre-built Python-based MCP server that automatically exposes Confluence tools to the agent. Watch the architecture in action as the Java-based agent seamlessly queries local functions (get_city_coordinates, get_weather) and hands the data off to the MCP tools to generate a Confluence page. We also break down the critical difference between running a local Python sidecar process for local testing versus connecting to dedicated, remote MCP URLs for high-concurrency production environments. Drop a comment below with your feedback on the delivery, and let me know how you are structuring your agent architectures. ⏱️ Chapters & Timeline 00:00 - The Problem with Manual Functions: Recapping our previous agent setup and why writing custom function tools doesn't scale. 00:29 - The Goal: Equipping the agent to automatically pull weather data and publish it directly to a Confluence page. 00:57 - Setting Up the MCP Server: Installing and configuring a Python package to act as our unofficial Atlassian MCP server. 02:18 - Verifying Tool Access: Checking the LLM's memory to confirm it automatically registered the new Confluence functions alongside our custom country/weather tools. 03:42 - What is Model Context Protocol? Understanding the open standard that handles method definitions and authentication for third-party platforms. 04:26 - Under the Hood: How the Java agent connects to the Python process, evaluates the prompt, and executes the correct tool chain. 05:53 - Live Demo: The Trip Planner agent successfully gathers Tokyo's weather and writes a detailed page in our Confluence "travel" space without explicit prompt engineering. 07:35 - Production Architecture: Why you should transition from local processes to remote, dedicated MCP servers for enterprise deployment. Slides: https://akhilojha.github.io/google-adk-tutorial-slides/section-3 #aiagents #javaai #aiprogramming #mcpserver #adk #googleadk