MCP (Model Context Protocol) in Claude Code — AI Glossary

MCP (Model Context Protocol) is an open standard that lets Claude Code connect to external tools, APIs, databases, and data sources through a structured interface. Instead of hardcoding integrations or relying on ad-hoc shell commands, MCP gives Claude Code a consistent way to discover and invoke external capabilities — databases, monitoring systems, version control APIs, or any custom tool you expose as an MCP server.

Why MCP Matters for Claude Code

MCP transforms Claude Code from a self-contained terminal agent into a programmable platform. Without MCP, Claude Code is limited to your local filesystem and shell commands. With MCP, it can query your production database, pull metrics from a monitoring dashboard, or call an internal API — all within the same agentic workflow.

This is significant for teams: you define once what tools Claude Code can access, and every developer gets the same capabilities without manual setup. Read how this fits into the broader extension model in Claude Code's Extension Stack: Skills, Hooks, Agents, and MCP, and see how MCP servers are built in How to Create an MCP Server.

How MCP Works in Claude Code

MCP operates on a client-server model. Claude Code acts as the MCP client; you run one or more MCP servers that expose specific capabilities. When Claude Code needs a tool, it queries the server via a defined protocol — the server responds with available tools and their schemas, and Claude Code calls them like any other function.

Key mechanics:

• Tool discovery: Claude Code asks an MCP server what tools it exposes and what parameters each accepts
• Structured calls: Invocations use a typed schema, not raw shell strings — reducing ambiguity and injection risk
• MCP server URL: You configure the server endpoint (local socket or remote URL) in Claude Code's settings; Claude Code connects at startup
• Multiple servers: A single Claude Code session can connect to several MCP servers simultaneously — one for your database, one for your CI system, one for your internal docs API

MCP is conceptually similar to a REST API in that both define structured endpoints for external capabilities. The difference is that MCP is optimized for LLM consumption: tool schemas are written to be interpreted by a model, not just a human developer. See the full picture of how this layer fits alongside skills and hooks in Claude Code's Seven Programmable Layers.

Related Terms

• Agentic Coding: The broader paradigm where AI agents like Claude Code plan and execute multi-step engineering tasks autonomously — MCP expands what those tasks can touch
• Agent SDK: The underlying SDK for building agent-based workflows; MCP servers can be built and registered using agent tooling patterns

---

Want more AI insights? Subscribe to LoreAI for daily briefings.