A language model is fundamentally a well-trained conversationalist. It has read an enormous amount of text. It can reason, formulate, and explain. But it lives in a sealed room. It can talk about SQL, but it can't run a database query. It can discuss a calendar, but it can't create a meeting.
MCP is the door out of that room.
What MCP is
The Model Context Protocol is an open protocol from Anthropic that standardises how AI agents communicate with external tools and systems.
Think of it as a shared language. Claude speaks one language. GitHub speaks another. Your database speaks a third. Without an interpreter in the middle, they can't understand each other. MCP is the interpreter: a standardised format for how calls look, how responses are structured, how verification works.
Claude Code — the tool you've been using in this bootcamp — is built entirely on MCP. Every time it reads a file from your project, that's an MCP call. Every terminal command it runs. Every GitHub operation. Everything happens via standardised MCP tools, not custom code for each service.
Three parts
MCP defines three types of capabilities an agent can have access to.
Tools are functions the agent can call: fetch a list of issues from GitHub, send an email, write to a database, run a search. Each tool has a name, a description, and a defined input and output.
Resources are data the agent can read: files, documentation, context windows with information about a project. Readable but not necessarily executable.
Prompts are reusable instructions: templates for common tasks that an agent can invoke with specific parameters instead of rewriting the instructions each time.
Why being a standard matters
Before MCP, every AI platform had its own format for tool calls. Integrating with Claude required one format. Integrating with GPT required another. Supporting both meant double the work.
MCP solves that. Anthropic, Google, OpenAI, Cursor, and GitHub Copilot all support the protocol. An MCP server that exposes your database works with all of these agents. You build once.
It's the same principle that allowed the web to grow. HTML is an open format all browsers can read. Nobody owns it. It lowered the bar for building something enormous.
In practice
Supabase has an official MCP server. Configure it once and Claude Code can read your tables, run queries, and understand your schema without you needing to paste schema information into every conversation.
There are MCP servers for GitHub, Slack, Google Drive, Notion, Stripe, and hundreds of other services. The ecosystem is growing fast. The agent you set up today can integrate with an ecosystem of services communicating via a shared protocol.