A practical account of migrating a multi-agent LangGraph system from scattered local tool definitions to a centralized MCP (Model Context Protocol) server. Covers building an MCP server with FastMCP, choosing between stdio and HTTP transports, integrating with LangGraph via langchain-mcp-adapters, moving human-in-the-loop approval logic to a single protocol boundary layer, and real production pitfalls like server crashes, wrong tool routing, and long-running approval gates. The migration eliminated tool duplication across seven agents and decoupled ML and application teams via a clean contract.
Table of contents
In This ArticleWhat is MCP?Building the MCP ServerStdio vs HTTPConnecting it to LangGraphHuman-in-the-Loop at the Protocol BoundaryWhat can break in Production and Why?Impact of MCP on our Agentic SystemConclusion777 Impressions