What is an MCP Server?

The problem it solves

AI language models are good at reasoning. What they are not good at, by design, is knowing things that exist outside their training data: your codebase, your database, your project decisions, your team's latest updates.

Before MCP, every team connecting an AI agent to an external tool had to build a custom integration. A function to call an API, some prompt engineering to format the result, error handling: all of it bespoke, all of it fragile. And if you wanted your agent to work across tools, you multiplied that effort by however many tools you had.

MCP standardizes this. It defines a shared protocol so that any AI client (Claude, GitHub Copilot, Cursor) can talk to any MCP server without custom glue code. Build the server once, and every compatible client can use it.

How it works

MCP follows a client-server architecture. The client is the AI tool you are working in: your IDE, your chat interface, your agent runtime. The server is an external process that exposes capabilities the client can call.

Those capabilities come in three forms:

• Tools: functions the agent can call to perform an action or retrieve data. Querying a database, fetching a document, calling an API.
• Resources: files or streams the client can read, similar to GET endpoints but accessible inline in the conversation.
• Prompts: pre-written templates the server exposes that the client can invoke as shortcuts.

When an agent needs information it does not have, it issues a tool call. The MCP server receives the request, runs whatever logic is necessary, and returns a result. The agent reads the result and continues reasoning. The whole exchange happens in the background: from the user's perspective, the agent just knows things.

The protocol is open

MCP was introduced by Anthropic in late 2024 and immediately released as an open standard. Anyone can build an MCP server or client. There is no licensing fee, no mandatory dependency on any specific AI provider.

This matters because it means the ecosystem compounds. A tool that exposes itself as an MCP server can be used by any current and future AI client that adopts the protocol, not just the one it was originally built for. GitHub Copilot, Claude Code, Cursor, Continue, Cline, and others already support it. The list keeps growing.

MCP for project knowledge

One of the most practical uses for an MCP server is exposing project-specific knowledge to an agent. Instead of keeping that knowledge in a system prompt (which gets stale and hits length limits), you put it in a server the agent can query on demand.

The agent asks for what it needs, when it needs it. Architectural decisions, domain concepts, operational constraints, recent changes: all queryable in real time, always up to date.

This is exactly what meridiaan does. Every project gets its own MCP endpoint. The topics you write become the tools the agent calls. You configure the endpoint once in your AI client and, from that point on, your agent has full access to the project knowledge base without you ever having to re-explain anything.