Module 06 — MCP (Model Context Protocol)

6.1 — What is MCP? The universal AI-to-tools protocol

Until now, every AI tool had to build its own integrations with every external service. One AI client, ten tools: ten specific integrations, each with its own format, authentication, and limitations. Three AI clients, ten tools: thirty integrations. MCP -- Model Context Protocol -- eliminates this complexity.

MCP is an open, standardized protocol that defines how an AI assistant communicates with external tools. Think USB-C: a single universal connector. Each tool is implemented once as an MCP server, and every compatible AI client understands it natively. You go from N x M integrations to N + M.

The architecture is built on a client-server model. Claude Code acts as the MCP client: it discovers the tools available on a server, understands their parameters, and calls them when relevant. The MCP server exposes three types of primitives:

• Tools: executable actions (create a Jira ticket, run a SQL query)
• Resources: consultable data (a database schema, a config)
• Prompts: reusable templates for common tasks

What makes MCP particularly efficient is dynamic discovery. Claude Code doesn't need to load every detail of every tool at startup. Thanks to the Tool Search mechanism, only tool names are loaded initially ("deferred" tools). Their full schema -- parameters, description -- is only fetched when Claude actually needs it. You can connect dozens of MCP servers without blowing up your context window.

The ecosystem is already rich: hundreds of community servers exist for databases, project management tools, cloud APIs, design services, and monitoring. And you can build your own MCP servers to expose your company's internal tools.