Skip to main content
Using Claude

What Is MCP (Model Context Protocol)? Architecture and Security

MCP (Model Context Protocol) is an open standard for how AI models connect to external tools. Covers the architecture, how it works, and security considerations.

By
🌐 This article was machine-translated and may contain inaccuracies. Read the Korean original if in doubt.

MCP (Model Context Protocol) is an open standard Anthropic released in November 2024. It standardizes how AI models connect to external tools, databases, and services.

🟢 Model references match the current lineup · model notice · Fable subscription
🟢 Model references match the current lineup · Claude Opus 5.5 / Claude Sonnet 5 / Claude Haiku 4.5 (higher tier: Claude Fable 5.1). This notice changes only when Anthropic ships a new model.

Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (

🟢 Model references match the current lineup · Claude Opus 5.5 / Claude Sonnet 5 / Claude Haiku 4.5 (higher tier: Claude Fable 5.1). This notice changes only when Anthropic ships a new model.
0/M input, $50/M output tokens). The one-time
🟢 Model references match the current lineup · Claude Opus 5.5 / Claude Sonnet 5 / Claude Haiku 4.5 (higher tier: Claude Fable 5.1). This notice changes only when Anthropic ships a new model.
00 credit applied only to the Fable 5 transition and is not offered for 5.1. Some coding and debugging requests may be answered by an Opus model due to a security classifier (both models). See the Fable 5.1 guide and the Fable 5 availability guide for details.

MCP: AI app ↔ MCP server ↔ external tools AI app (Host) Claude · Cursor · VS Code MCP server External tools & data Drive · Slack · GitHub · DB MCP · JSON-RPC Tools (run) Resources (read) Prompts (template) An open standard connecting N apps + M services at N+M cost · launched Nov 2024

The problem MCP solves

Before MCP, you had to write separate integration code for every LLM and every tool (the N×M problem). For example, you'd build a GitHub integration for Claude and another for ChatGPT separately. MCP lets you expose a tool once as an MCP server, and any host that supports MCP (Claude Desktop, Cursor, VS Code, etc.) can use it.

This standard is no longer Anthropic-only. After being open-sourced in November 2024, it was adopted by major AI providers including OpenAI (2025) and Google, and is supported by developer tools such as VS Code and Cursor. MCP is now maintained by the community under open governance within the Linux Foundation.

Architecture

  • Host: the AI app the user interacts with (Claude Desktop, an IDE, etc.)
  • Client: discovers servers within the host and routes tool calls
  • Server: exposes access to actual tools and data (GitHub, PostgreSQL, Slack, etc.)

Communication is based on JSON-RPC 2.0. Function calling is the underlying mechanism that actually executes tools, and MCP provides a standardized connection layer on top of it.

Why it matters

MCP is vendor-neutral. Build a tool integration once and you don't have to rewrite it when you switch models from Claude → GPT → Gemini. Integration investment is decoupled from model choice. Major providers like OpenAI and Google have also adopted MCP.

Security considerations

An MCP server gives the AI model direct access to tools. A misconfigured server could let the AI read data it shouldn't, or write or delete data. Follow these practices:

  • Restrict to read-only scope where possible
  • Run the server under a least-privilege service account
  • Enable client-side execution confirmation for destructive actions (delete/modify)

Note that MCP's security model recommends explicit user consent for tool calls, but actual enforcement is the responsibility of the host app. Check how your host implements consent.

Why MCP: ad-hoc links → one standardWithout MCP AI appAI appAI appToolToolTool 앱×Tool마다 제각각 (N×M) With MCP AI appAI appAI appToolToolTool MCP Simplified by one standard Conceptual · source modelcontextprotocol.io · as of 2026.6

Token cost

Each MCP tool definition and call result is included in the context, so it consumes tokens. Connecting many servers at once can increase token usage, so it's best to enable only the servers you need.

For which servers are worth connecting, see recommended MCP servers.

To see which servers exist and where to get them, check where to find MCP servers.

MCP is a fast-evolving standard. Check the official docs (modelcontextprotocol.io) for the latest spec.

See also: Claude Code MCP setup

Related MCP guides: How to connect MCP · MCP server troubleshooting · MCP for beginners · MCP connector security

Was this helpful?

Keep reading

Connect Blender to Claude Code with an MCP Server: Step-by-Step Guide

Learn how to control Blender's 3D viewport from Claude Code using an MCP server, from installing uv to registering the server with claude mcp add.

55

GPT-5.6 Sol's Over-Verification Backlash — and How to Stop It in Claude Code

GPT-5.6 Sol drew criticism for re-running entire test suites on tiny changes and burning through usage limits. Here's how Claude Code users can guard against the same failure mode with an open-source plugin plus Claude Code's own Plan Mode and Hooks.

33

Gave Up Installing ComfyUI? Use It From Claude Instead (Comfy MCP)

Why people give up on ComfyUI (path issues, GPU limits, 15GB, node graphs) and how Comfy MCP lets you drive it from Claude instead. Which connection to pick and what it costs, from official docs as of August 2026.

86

Setting Up the Sentry MCP Server in Claude Code: From Error Logs to Fixes

Connect the Sentry MCP server to Claude Code and query issues, read stack traces, and start fixing bugs without leaving your terminal. This guide covers OAuth setup, the claude mcp add command, and scoping tokens safely.

19