> Graphify: the code knowledge graph for agents. Canonical: https://graphify.com/blog/what-is-an-mcp-server-a-developers-guide · Full index: https://graphify.com/llms-full.txt

# What is an MCP server? A developer's guide

*April 20, 2026 · Syed Fahad*

An MCP server is a program that exposes tools, data, or prompts to an AI application through the Model Context Protocol. The AI app, called the host, connects to the server through a client. The model can then call the server's tools, such as querying a database or a code graph, without custom glue code for each app. Graphify, an open-source CLI, can run one that serves a code graph to your agent.

## What is an MCP server, in plain terms?

An MCP server is an adapter between a model and a system. It describes what it offers in a standard format, and any MCP-capable host can use it. The [Model Context Protocol site](https://modelcontextprotocol.io/introduction) compares MCP to a USB-C port for AI applications: one connector, many devices.

Before MCP, each assistant needed its own plugin format. With MCP, one server can work in several hosts, such as Claude Code, Cursor, and other clients that implement the protocol.

## What can an MCP server expose?

The protocol defines three main building blocks. The [server concepts page](https://modelcontextprotocol.io/docs/learn/server-concepts) describes them.

| Building block | What it is | Who decides to use it |
|---|---|---|
| Tools | Functions the model can call, with typed inputs | The model |
| Resources | Read-only data, such as files or records | The application |
| Prompts | Reusable templates for common tasks | The user |

Most developer servers start with tools. A tool has a name, a description, and a JSON schema for its arguments. The model reads the description and decides when to call it.

## How does a host talk to an MCP server?

Messages use JSON-RPC 2.0. The [specification](https://modelcontextprotocol.io/specification) defines two standard transports. With stdio, the host starts the server as a local process and talks over standard input and output. With Streamable HTTP, the server runs as a web service and can serve many clients.

Local developer tools usually use stdio. Nothing listens on a network port, and the server runs with your user's permissions.

## When should you build or add an MCP server?

Add one when a model needs repeated, structured access to something. A single shell command works for a one-off check. A server is better when the model will call the same capability many times in a session, and you want typed inputs and short, predictable outputs.

Keep tool outputs small. Every result goes into the model's context window. A tool that returns a whole file defeats the point.

## What does an MCP server for a code graph look like?

Graphify is an open-source CLI that builds a code knowledge graph from a repository. It can also serve that graph over MCP on stdio. Install the extra with `uv tool install "graphifyy[mcp]"` or `pip install "graphifyy[mcp]"`, build a graph, then start the server:

```bash
python -m graphify.serve graphify-out/graph.json
```

The server exposes four tools: `query_graph`, `get_node`, `get_neighbors`, and `shortest_path`. An agent can ask for the neighbors of a function or the path between two classes. It gets a few lines of structure back, instead of reading files to find them.

To register it, add the command to your host's MCP configuration. Each host documents its own file location, so follow the host's MCP guide.

## What are the security basics?

An MCP server can do whatever its code allows. Run servers you trust, from sources you can read. Prefer read-only tools where possible. Check what a server sends over the network. The Graphify server reads a local `graph.json` and has no telemetry.

## What is the difference between an MCP host, client and server?

The host is the AI application, such as Claude Code or an IDE. Inside the host, one MCP client keeps a connection to each server. The server is the program that exposes tools, resources and prompts. One host can run many clients, each talking to a different server.

The server can run locally over stdio or remotely over HTTP. The host decides which servers to start and which tools the model may call.

## FAQ

### Is MCP only for Claude?

No. Anthropic introduced it as an open protocol, and other hosts implement it. Check your tool's documentation for MCP support.

### What language do I write an MCP server in?

Any language that can speak JSON-RPC. Official SDKs exist for several languages, including Python and TypeScript.

### Is an MCP server the same as an API?

Not exactly. It wraps capabilities in a standard format that models can discover. It often calls an API underneath.

### Do I need MCP to use Graphify?

No. Assistants can read `GRAPH_REPORT.md` and run `graphify query` in a shell. MCP is optional for repeated tool access.

### Does a stdio server open a network port?

No. The host starts it as a child process and talks over standard input and output.

## Try it

Graphify is open source on [GitHub](https://github.com/Graphify-Labs/graphify). Install it with [uv](https://docs.astral.sh/uv/): `uv tool install graphifyy`, run `graphify install`, and build your first graph. The [MCP server glossary entry](https://graphify.com/glossary/mcp-server) goes further. For keeping tool output small, see [what fills the Claude Code context window](/blog/claude-code-context-window-what-fills-it-and-how-to-keep-it-lean). See also [measuring token spend per task](/blog/measuring-token-spend-per-task-in-claude-code). Setup details are in the [installation docs](https://docs.graphify.com/open-source/installation).