Working with Anthropic’s Model Context Protocol (MCP) — Part 1

Building a simple MCP Server & Client using Server-Send Events on Google Colab

Introduction

I started reading about the Model Context Protocol (MCP) from Anthropic a few months ago after being introduced to it by a coworker. Anthropic first announced the MCP on November 24, 2024. There was clearly a lot of excitement in the AI engineering community if I judged purely on the number of articles and videos that were published that month on the subject.

What excited me most at the time was how the protocol stripped away the framework specific dependencies of the very daunting and growing list of agentic application and orchestration frameworks — LangGraph, CrewAI, Bedrock, PydanticAI, etc. The MCP offers up a standard that the community at large can use (and contribute to) to integrate their agentic workflows with one another based on varying scopes or intent (e.g. data access or resources only, purpose built tools, etc.) without getting bogged down in the stylistic, business driven, or architectural decisions made by any one group or organization.

In short, the MCP is an open protocol that standardizes how agentic workflows and applications communicate and share context, resources, tools, prompt catalogs and more with one another.

The image below emphasizes some of the myriad of ways an MCP client might engage with a data resource using the SSE…