What is MCP? The Model Context Protocol explained
MCP (Model Context Protocol) is an open standard for connecting AI applications such as Claude or ChatGPT to external tools and data. An MCP server exposes capabilities, for example a database lookup or a CRM search, and any MCP-compatible application can discover and use them without a custom integration for each pairing. For a business, it means you build one connector per system and reuse it across AI tools.

The problem MCP solves
A language model knows what it was trained on and what is in the conversation. It does not know your order table, your CRM or last week’s tickets. To be useful inside a business, an AI application has to reach those systems.
Before MCP, every connection was custom. Each AI application had its own plugin format, so connecting five applications to ten systems meant up to fifty integrations. MCP replaces that with one protocol. A system is wrapped once as an MCP server, and every application that speaks MCP can use it.
Anthropic introduced MCP as an open standard in November 2024. The specification and SDKs are public at modelcontextprotocol.io, and the protocol is supported by AI applications from several vendors, including Claude, ChatGPT, VS Code and Cursor.
The three participants: host, client, server
MCP uses a client-server architecture with three roles. The definitions below follow the official architecture documentation.
- MCP host. The AI application the user works in, such as Claude Desktop, Claude Code or VS Code. The host coordinates one or more MCP clients.
- MCP client. A component inside the host that keeps a connection to one MCP server and obtains context from it. The host creates one client per server.
- MCP server. A program that provides context to MCP clients. It can run locally or remotely.
In practice: a user asks Claude a question about an order. Claude (the host) has a client connected to your orders server. The model decides to call the server’s order lookup tool, the client sends the request, the server runs the lookup and returns the result, and the model writes the answer.
The word “server” describes the role, not the location. A local server runs as a process on the user’s machine. A remote server runs on a platform and serves many clients.
What a server can expose: tools, resources and prompts
MCP defines three core primitives that servers can expose.
| Primitive | What it is | Example |
|---|---|---|
| Tools | Executable functions the AI application can invoke to perform actions | Look up an order, search the CRM, run an approved report |
| Resources | Data sources that provide contextual information | A file’s contents, a database schema, a record |
| Prompts | Reusable templates that structure an interaction | A standard “summarize this account” prompt |
Each primitive has methods for discovery and use. A client calls tools/list to learn what tools exist, and tools/call to run one. Each tool has a name, a description and an input schema written in JSON Schema, which tells the model what arguments the tool accepts. Because listing is dynamic, a server can show different tools to different callers.
Clients can expose primitives too. The main one is elicitation, which lets a server ask the user for more information or for confirmation of an action. Two older client primitives, sampling and logging, are deprecated as of protocol version 2026-07-28, so new servers should not depend on them.
How the messages travel: the two layers
MCP has two layers.
The data layer defines the messages. It is based on JSON-RPC 2.0, a simple request and response format. This layer covers discovery, the primitives above and notifications, which a server sends when something changes, for example when its tool list is updated.
The transport layer defines how messages move. MCP supports two transports:
- Stdio. The host launches the server as a local process and talks to it over standard input and output. There is no network involved. This suits tools that run on a developer’s or employee’s own machine.
- Streamable HTTP. The client sends messages by HTTP POST, and the server can optionally stream responses using Server-Sent Events. This is the transport for remote servers. It supports standard HTTP authentication, and the documentation recommends OAuth for obtaining tokens.
The same JSON-RPC messages run over both transports, so a server’s logic does not change when it moves from local to remote.
One detail for technical readers: in the current revision of the protocol (2026-07-28), MCP is stateless. Every request carries the protocol version and the client’s capabilities, so a server can process each request on its own, and clients can query a server’s supported versions and capabilities with a server/discover request. Stateless servers are easier to scale horizontally behind a load balancer.
MCP is not a replacement for your APIs
An MCP server usually sits in front of an existing API or database. The API still does the work. The MCP server describes that work in a form a model can discover and call. The full comparison is in MCP vs API.
Security considerations
MCP standardizes the connection. It does not decide what the connection may do. That is the job of whoever builds and runs the server. These are the points to settle before a server touches real data.
What the server can do. Every tool is a capability handed to a model. Start read-only. Add write tools one at a time, with limits. A tool that accepts a raw SQL string gives the model your whole database. A tool that runs one of a fixed list of approved queries does not.
Whose permissions apply. If the server connects to your CRM with one admin account, every user of the AI application inherits admin access through it. The server should act with the permissions of the person asking.
What the model sees. Tool results go into the model’s context and may be sent to a model provider. Mask personal data that the task does not need.
Content that hijacks the model. Text returned by a tool (an email body, a ticket, a web page) can contain instructions aimed at the model. This is prompt injection. Treat all tool output as untrusted data, keep permissions enforced on the server, and require confirmation for sensitive actions.
Reconstructing what happened. Log every tool call with the user, the arguments and the result. Without a log, you cannot investigate an incident or prove compliance.
Who may connect. Remote servers need authentication. Local HTTP servers need protection against requests from malicious web pages, which is why the official SDK ships Host and Origin header validation.
Trust in the server itself. Installing a third-party MCP server is running third-party code with access to your AI application. Review it as you would any dependency.
Shai’s on-prem MCP gateway applies these rules as defaults: read-only, approved queries only so the model never writes SQL, per-user permissions, PII masking, an audit log and prompt-injection defenses. The build details are in how to build an MCP server.
How enterprises use MCP
Three patterns are common.
Internal data access. Employees ask questions in Claude or ChatGPT and get answers from the ERP, CRM, SQL databases and Microsoft 365, with their own permissions applied. This is what Shai’s gateway does. It installs at the customer with Docker Compose, has 13 connector types, and keeps the data on the customer’s side.
Agent tooling. A company building its own AI agents writes the system connections once as MCP servers and reuses them across agents. This cuts the integration cost that dominates most agent budgets, as described in how much it costs to build an AI agent.
Developer tooling. Engineering teams connect coding agents such as Claude Code to issue trackers, error monitoring and internal docs, so the agent can read a bug report and the related logs before it writes a fix.
A gateway in front of many servers gives one place to authenticate users, apply permissions, mask data and keep the audit log. Without one, each server solves those problems separately.
When MCP is the wrong tool
- One application calling one API. Direct function calling is simpler.
- Batch jobs and system-to-system data sync. No model is choosing tools, so use the API.
- High-volume, latency-sensitive paths where the steps are fixed in code.
MCP earns its place when a model chooses the tools, and when more than one AI application needs the same systems.
What to do next
List the systems your team would query from an AI assistant if it were safe to do so, and mark which contain personal or financial data. That list is the scope of an MCP project. Our AI integration service builds custom MCP servers and deploys the on-prem gateway. Build projects start at $25,000 with a fixed price per phase, and the first one-hour consultation is free.