What is MCP (Model Context Protocol)?
An open standard from Anthropic that lets LLMs connect to tools, databases, and apps through a single unified protocol.
MCP (Model Context Protocol) is an open standard released by Anthropic in late 2024 that lets LLMs connect to tools, databases, and apps through a unified protocol — avoiding the need for custom integrations for every model × tool pair.
Why MCP?
Before MCP:
- Every LLM (Claude, GPT, Gemini) had its own function-calling syntax
- Every tool (Slack, GitHub, Notion) needed its own integration code
- Number of integrations =
N models × M tools(a lot!)
With MCP:
- A tool implements a single MCP server → any LLM that supports MCP can use it
- Number of integrations:
N + Minstead ofN × M
Roles
┌─────────────┐ MCP protocol ┌─────────────┐
│ MCP Client │ ←──────────────────→ │ MCP Server │
│ (LLM apps) │ (tool/data calls) │ (tool side) │
└─────────────┘ └─────────────┘
- Client: app that calls the LLM (Claude Desktop, Cursor, Claude Code, …)
- Server: app that exposes capabilities (Slack server, Postgres server, Filesystem server)
Available MCP servers (2026)
- Filesystem: read/write local files
- Postgres / SQLite: query a database
- GitHub: create issues, read PRs, search code
- Slack / Discord: send/read messages
- Brave Search: search the web
- Puppeteer: drive a browser
→ One install command, and Claude Code can use it instantly.
When should you use MCP?
✅ Building an agent with many tools — no need to code a wrapper for each ✅ You want to share tool integrations across the team ✅ You use multiple LLM clients (Claude Desktop at home, Cursor at work)
❌ You only need one simple tool → plain function calling is usually enough ❌ Highly custom internal tooling → it may be faster to code direct integration
Quick setup (Claude Desktop)
claude_desktop_config.json:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/Documents"]
}
}
}
Restart Claude → Claude can now read/write files in /Users/me/Documents.
Adoption
By mid-2026: Claude, OpenAI, and a number of IDEs and apps support MCP. It’s gradually becoming the de facto standard for tool use.