Last month, a simple database schema update exposed a messy reality: tool definitions scattered across three different locations in our agentic system. Each agent maintained its own version of similar tools, with approval logic wired directly into graph edges. A schema change required touching four files and re-testing each agent separately, hoping nothing downstream would break silently.

MCP (Model Context Protocol), published by Anthropic in late 2024, solved this by standardizing tool discovery and calling. Instead of defining tools locally within each orchestrator, we now run them on a separate server that agents connect to at runtime. This creates an interoperability boundary—any MCP-compatible client works without custom integration, whether LangGraph today or a different framework tomorrow.

Building the server uses FastMCP, which handles schema generation from type hints automatically. The key gotcha: never write to stdout since MCP uses it for communication. Teams can now own separate domains cleanly—the ML team owns tools while the application team owns the graph—without sharing codebases or stepping on each other's work.

Production deployment requires choosing between stdio for local development and streamable HTTP for containerized services. The protocol's transport-agnostic design means compliance teams get a clean story for data residency requirements, with servers running anywhere without changing the interface.