MCP servers: giving AI models hands

If you've used Claude, ChatGPT, or any modern LLM recently, you've probably hit the wall: the model can reason about things, but it can't do things. It can't query your database, check your calendar, or look up the current weather without help. The Model Context Protocol (MCP) is Anthropic's answer to this problem — an open standard for connecting AI models to external tools and data sources.

What MCP actually is

MCP is a JSON-RPC based protocol that defines how an AI application (the "client") communicates with external services (the "servers"). Think of it like USB for AI tools — a standardized way to plug capabilities into any model that supports the protocol.

An MCP server exposes three types of primitives: tools (functions the model can call, like querying a database), resources (data the model can read, like file contents), and prompts (reusable templates). The client — typically an AI chat application — discovers what the server offers and makes those capabilities available to the model during conversation.

Why it matters

Before MCP, every integration was bespoke. If you wanted Claude to query your PostgreSQL database, you'd write custom code. If you also wanted it to search your company's Confluence, that's another custom integration. Multiply this by every tool a team uses and it becomes unmanageable.

MCP servers are reusable. Someone writes a PostgreSQL MCP server once, publishes it, and anyone can connect it to any MCP-compatible client. There are already open-source servers for databases, GitHub, Slack, Google Drive, file systems, web scraping, and dozens of other services. The ecosystem is growing rapidly.

Building one is straightforward

An MCP server is a lightweight process — typically a Node.js or Python script. The official SDKs handle the protocol negotiation, so you're mostly writing tool definitions and their handler functions. A minimal server that provides a single tool might be 50 lines of code.

At UpperBlue, we built an MCP server that exposes Icelandic open data — earthquake data from the Met Office, borehole registries, air quality measurements, and calendar information. It connects our NestJS API to any MCP-compatible client, meaning Claude can answer questions about Icelandic seismic activity, sunrise times, or air quality directly in a conversation.

The practical implications

For businesses, MCP changes what's feasible with AI assistants. Instead of a chatbot that only knows what's in its training data, you can have one that queries your live inventory, checks your CRM, reads your internal documentation, and creates Jira tickets — all through standardized, auditable connections.

The security model is straightforward: MCP servers run in your infrastructure, you control what data they expose, and the transport layer (typically stdio or SSE) keeps communication local. The model never gets direct database access; it calls defined functions with validated parameters.

Where this is heading

MCP is still young — Anthropic open-sourced it in late 2024 — but adoption is fast. Claude Desktop, Cursor, Windsurf, and several other AI tools support it natively. The protocol spec is evolving to support authentication, streaming, and more complex interaction patterns.

The bet is that a standard protocol layer between AI and tools will do for AI applications what HTTP did for the web: create a common interface that enables an ecosystem. If the bet pays off, the value of an AI system will increasingly be defined by the quality of its MCP connections — the data and tools it can actually access — rather than just the model's raw capabilities.