What is an MCP server?
MCP in a nutshell
MCP stands for Model Context Protocol. It's a specification that defines how AI applications (called hosts) can connect to external capabilities (called servers) in a standardized way. Think of it like USB-C for AI tools: instead of building custom integrations for every model and every tool, you build one MCP server that any compatible AI host can use.
An MCP server can provide three main types of capabilities: tools (functions the model can call), resources (data the model can read), and prompts (templates for interactions). The server communicates with the host over a transport like stdio or HTTP with Server-Sent Events (SSE).
- Tools: executable functions (e.g., send email, query database)
- Resources: file-like data (e.g., documents, API responses)
- Prompts: reusable templates for common tasks
- Transports: stdio (local) or HTTP+SSE (remote)
Why it matters
Before MCP, connecting an AI assistant to a new tool meant writing custom code for each host (Claude Desktop, an IDE plugin, a custom chatbot). MCP standardizes that interface, so a single server works across any MCP-compatible client.
This reduces duplication and makes it easier to share and reuse integrations. For example, a GitHub MCP server can be used by Claude Desktop, Cursor, and other clients without modification. The protocol is open source and has SDKs in Python, TypeScript, and other languages.
Common examples
Popular MCP servers include those for file systems, databases (PostgreSQL, SQLite), web browsing (Puppeteer), and SaaS APIs (GitHub, Slack, Google Drive). Many are community-built and available on GitHub or package registries.
You can also build your own MCP server to expose proprietary data or internal tools. The protocol is designed to be simple enough for a single developer to implement in an afternoon.
- Filesystem server: read/write local files
- Database server: run SQL queries
- GitHub server: manage issues, PRs, repositories
- Web browser server: fetch and interact with web pages
Common mistakes
- Confusing MCP with a specific AI model or vendor; it's an open protocol, not a product.
- Thinking MCP servers are only for Anthropic's Claude; any client can implement MCP.
- Assuming MCP replaces function calling; it's a standardized way to expose functions, but the underlying model still uses function calling to invoke them.