What is Claude MCP? Model Context Protocol

The Problem MCP Solves: N×M Integration Hell

Before MCP, connecting an AI Tool to a data source was a tangled affair — every vendor had to write its own integration:

• 5 AI Tools (Claude, ChatGPT, Cursor, Copilot, Gemini)
• 10 data sources (GitHub, Slack, Notion, Postgres, Google Drive, etc.)
• Required 5 × 10 = 50 integrations for every pair to work together

The more AI Tools and data sources we have, the more integrations grow as N×M — not scalable, full of duplicate code, bugs, and security risks. Anthropic created MCP as a shared standard — write the MCP server once, use it with any AI that supports MCP. This reduces the equation from N×M to N+M.

Architecture: Client-Server over JSON-RPC 2.0

MCP uses a Client-Server architecture communicating over JSON-RPC 2.0 — the same protocol family as the Language Server Protocol (LSP) that VS Code and other IDEs use to connect with language servers.

Role	Responsibility	Examples
MCP Host	The application the user opens — has the LLM inside	Claude Desktop, Claude Code, Cursor, thClaws
MCP Client	Host's representative that talks to each Server (1 client per server)	Lives inside the Host — not directly visible to the user
MCP Server	An application that exposes tools/resources/prompts to be called by Clients	GitHub MCP server, Slack MCP server, Filesystem MCP server

The 3 Capabilities an MCP Server Exposes

Each MCP Server can expose 3 types of capability for the Host to use:

Capability	What it is	Examples
Tools	Functions the AI can call — may have side effects (modify data, send messages)	create_issue, send_slack, query_db, run_command
Resources	Data the AI can read — read-only, no side effects	file://path, postgres://table, gdrive://doc
Prompts	Pre-built prompt templates the Host can pick	"summarize_pr", "explain_query", "code_review"

Handshake & Discovery — Getting Started

When the Host starts up (e.g. Claude Desktop boots), it talks to the configured MCP Servers in this sequence:

1. Initialize — Client sends an initialize request with version and supported capabilities
2. Server Response — Server replies with its own version and capabilities
3. List — Client requests the list of tools/resources/prompts via tools/list, resources/list, prompts/list
4. Register — Host stores these into memory and offers them to the LLM during the conversation
5. Invoke — when the LLM decides to call a tool, the Client sends tools/call to the Server and receives the result back

Transport Modes — How Clients Talk to Servers

MCP supports 3 ways for Clients and Servers to communicate, each suited to different environments:

Transport	Best For	Detail
stdio	Local server on the same machine	Host spawns the Server's process and talks via stdin/stdout — simplest and most secure
HTTP + SSE	Remote server (legacy)	HTTP POST for requests, Server-Sent Events for streaming — older approach; Streamable HTTP is now recommended
Streamable HTTP	Remote server (modern)	A single HTTP endpoint supporting both request/response and streaming — the new recommended standard

MCP Servers Available Today

There are now over 10,000 MCP servers in the ecosystem — from Anthropic, vendor-official, and community sources. Popular categories:

Category	Example MCP Servers
Code & DevOps	GitHub, GitLab, Git, Docker, Kubernetes, Sentry
Communication	Slack, Discord, Email (Gmail/IMAP), Teams
Data	Postgres, MySQL, MongoDB, Redis, BigQuery, Snowflake
Productivity	Google Drive, Google Calendar, Notion, Linear, Jira
Browser & Web	Puppeteer, Playwright, Brave Search, Fetch
Filesystem	Local filesystem, S3, GCS, Azure Blob

Example: Configuring an MCP Server in Claude Desktop

Adding an MCP server in Claude Desktop is done via a config file at ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_xxxxxxxxxxxx"
      }
    },
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem",
               "/Users/me/projects"]
    }
  }
}

When Claude Desktop starts, it spawns both servers via stdio and begins the handshake automatically. The user can immediately tell Claude things like "create an issue on GitHub repo X" or "read a file in /Users/me/projects/foo".

Adoption in 2026 — The AI Industry Standard

MCP is no longer just an Anthropic project — it has become a de facto industry standard adopted by every major player:

• 97 million downloads/month of the MCP SDK across the ecosystem (March 2026)
• 10,000+ public MCP servers available
• December 2025: Anthropic donated MCP to the Linux Foundation under the Agentic AI Foundation — co-founded by Anthropic, Block, and OpenAI, with support from Google, Microsoft, and AWS
• Adopted by: Claude (Anthropic), ChatGPT (OpenAI), Gemini (Google), Cursor, Sourcegraph Amp, Factory, GitHub Copilot, thClaws

MCP vs Claude Skills — How Are They Different?

MCP and Claude Skills are complementary mechanisms — they are not replacements for one another:

Topic	MCP	Skills
Purpose	Connects AI to external tools/data sources	Bundles workflows + instructions + scripts for specialized tasks
Form	A separate Server process speaking JSON-RPC	A folder containing SKILL.md + scripts + assets
Triggering	LLM calls a tool by name and schema	Auto-trigger via whenToUse, or by slash command
Used Together	A Skill can call tools from MCP servers internally	A higher-level layer above MCP — encapsulating workflows that use MCP tools

Summary — Claude MCP in One Paragraph

Topic	Summary
What it is	An open standard from Anthropic for plug-and-play connection of AI Tools to external data/tool sources
Analogy	The USB-C of AI — one plug, every device
Technical	Client-Server architecture over JSON-RPC 2.0 (like LSP) — supports stdio, HTTP+SSE, and Streamable HTTP
3 Capabilities	Tools (callable) / Resources (readable) / Prompts (preset templates)
Ecosystem	97M downloads/month, 10K+ servers, supported by Anthropic/OpenAI/Google/Cursor/Microsoft/AWS
Governance	Linux Foundation (Agentic AI Foundation) since December 2025

MCP changed AI Integration from N×M to N+M — write a tool once and use it everywhere. It is the single standard that lets the AI Agent ecosystem grow exponentially in 2026.

- Saeree ERP Team

Read More — Related Articles

• Claude Skills: What are Agent Skills — A Complete Guide
• thClaws: A Thai-Made AI Agent Workspace That Supports MCP
• Source Leak: Claude Code Source Code Leak: 512,000 Lines

References

• Anthropic — Introducing the Model Context Protocol
• modelcontextprotocol.io — Official Documentation
• GitHub — Model Context Protocol Organization
• Descope — What Is the Model Context Protocol (MCP) and How It Works
• Truto — What is MCP? The 2026 Guide