MCP server: what it is and how it works
Jaseem Abbas
- Published
What is an MCP server?
An MCP server is a program that offers tools, data and prompts to AI applications in a standard way. The Model Context Protocol introduction (read 16 November 2026) calls MCP "an open-source standard for connecting AI applications to external systems", and compares it to a USB-C port: one shape of plug that many devices share.
Picture Ravi, a developer at a Bengaluru logistics firm. He wants his team's AI assistant to look up where a shipment is. He wraps the shipment database in a small MCP server with one tool, "find shipment". Now any AI application that speaks MCP can list that tool and call it, without Ravi writing a separate connector for each one. The firm and Ravi are made up for this example.
What are the parts of MCP: host, client, server, tools, resources and prompts?
MCP has three roles and three kinds of things a server can offer. The specification (revision 2026-07-28) names the roles: hosts are the AI applications that start connections, clients are connectors inside the host, and servers are the services that provide context and capabilities. Messages between them use JSON-RPC 2.0.
Feature: Tools
- What it is
- Functions for the AI model to execute
- Who it is for
- The model
Feature: Resources
- What it is
- Context and data
- Who it is for
- The user or the model
Feature: Prompts
- What it is
- Templated messages and workflows
- Who it is for
- The user
The tools page of the specification says tools let models interact with outside systems, "such as querying databases, calling APIs, or performing computations", and that each tool has a unique name and a described schema. It also says there should always be a human in the loop with the ability to deny tool calls.
How does an MCP server connect: stdio or Streamable HTTP?
An MCP server connects over one of two standard transports, called stdio and Streamable HTTP. The specification's transports page describes both:
- stdio: the client starts the server as a subprocess on the same machine and they swap messages over its standard streams. It suits a tool that runs on your own laptop.
- Streamable HTTP: the server exposes one HTTP address, the client sends each message to it as a POST, and the reply comes back as a single JSON object or a stream. It suits a server that runs on the web for many users.
For a server on the web, the specification says servers should implement proper authentication for all connections, so decide who may call what before you publish one.
How do voice agents use MCP servers?
A voice agent uses an MCP server the same way any AI application does: it lists the server's tools and calls one when the caller's request needs it. This is function calling with a shared plug. A function lives in one application's code, while an MCP server can offer its tools to any agent that connects. For the longer explanation, read what an MCP server is, for voice agents.
Bananaflow works with MCP in two directions:
- Your MCP servers, during a call. An agent can call your own MCP servers while it is on the phone, for example to look up an order. Read how Bananaflow agents use your own MCP servers during a call.
- Bananaflow's MCP server, while you build. Bananaflow has an MCP server of its own, so an AI assistant such as Claude Code, Codex, Claude Desktop or Cursor can create, edit, test and evaluate voice agents and their tools. Read how to build a voice agent with Claude Code and MCP.
How is an MCP server different from a webhook or an API?
An API is a way for one program to call another; an MCP server is an API shaped so an AI application can find its tools and call them without custom code for each one. A webhook works the other way round: a system sends data to you when something happens, such as a call ending. An MCP server waits to be asked.
FAQ
Frequently asked questions
What does MCP stand for?
MCP stands for Model Context Protocol. It is an open-source standard for connecting AI applications to outside systems such as files, databases, search tools and workflows.
What is an MCP server?
An MCP server is a program that offers tools, data and prompts to AI applications through MCP. An AI application connects to it, lists what it offers and calls its tools when a task needs them.
What is the difference between an MCP client and an MCP server?
The client sits inside the AI application, which is called the host, and starts the connection. The server is the service on the other end that provides the tools and context. Each client talks to one server.
What are tools, resources and prompts?
They are the three things a server can offer. Tools are functions for the AI model to run, resources are context and data, and prompts are templated messages and workflows for the user.
What is the difference between stdio and Streamable HTTP?
With stdio, the client starts the server as a subprocess and they talk over its standard streams. With Streamable HTTP, the server has one web address and the client sends each message to it as an HTTP POST.
Is an MCP server the same as function calling?
No. Function calling is what the model does when it asks for a tool to be run. An MCP server is one place those tools can live, with a shared way to list and call them, so many agents can use the same tools.
Can a voice agent use an MCP server during a call?
On Bananaflow, yes. An agent can call your own MCP servers during a call, for example to look up an order, next to HTTP API tools, transfer to a person or another agent, and end call.
Can an AI assistant build voice agents through MCP?
Yes. Bananaflow has an MCP server that lets Claude Code, Codex, Claude Desktop or Cursor create, edit, test and evaluate voice agents and their tools while you describe what you want.
