What Is MCP? Model Context Protocol Explained Simply

What MCP is: the simplest explanation

Model Context Protocol (MCP) is a standard way for AI models to connect to external tools and data sources. Think of it as a universal adapter — like USB-C for AI.

Before MCP, every AI tool that wanted to connect to an external service (a database, a file system, a project management tool, a documentation platform) needed a custom integration. If you wanted Claude to access your company's Jira board, someone had to build a specific Jira-to-Claude connector. If you also wanted it to access your Notion workspace, that required a separate Notion-to-Claude connector. And if you switched from Claude to ChatGPT, you needed entirely new connectors. Every combination of AI model and external tool required its own custom bridge.

MCP eliminates this by establishing a common protocol — a shared language — that any AI model and any external tool can use to communicate. Instead of building custom connectors for every combination, tool providers build one MCP server, and AI providers build one MCP client. Any MCP client can talk to any MCP server. The combinatorial explosion of custom integrations collapses into a simple, scalable ecosystem.

The real-world impact is that AI models stop being isolated. Without MCP, an AI assistant can only work with what you paste into the conversation or upload as a file. With MCP, it can reach into your tools — pulling data from databases, reading documents from your knowledge base, checking task statuses in your project management system — as a natural part of the conversation. The AI goes from being a clever chat partner to being a participant in your actual workflow.

MCP was originally developed by Anthropic and released as an open standard in late 2024. It has since been adopted by multiple AI providers and tool vendors, creating an expanding ecosystem of interoperable connections. For a broader look at the AI terminology landscape, the [AI Glossary](/glossary) provides definitions for hundreds of terms including [Model Context Protocol](/glossary/model-context-protocol).

Why MCP matters: the problem it solves

To understand why MCP matters, consider the fundamental limitation of today's AI tools. When you ask ChatGPT or Claude a question, the model can only work with two sources of information: its training data (which has a cutoff date and is necessarily general) and whatever you provide in the conversation (pasted text, uploaded files, or typed context).

This means that every time you want an AI to work with your specific data — your company's database, your project's codebase, your team's documentation — you must manually extract that data and feed it into the conversation. This is tedious, error-prone, and often impractical. You cannot paste a 50-million-row database into a chat window. You cannot upload your entire company wiki. You cannot keep the AI updated on real-time changes to your project management board.

MCP solves this by letting AI models access external data sources directly, in real time, as part of the conversation. Instead of manually pasting your latest sales data, you ask the AI "what were our top-performing products last quarter?" and it queries your database through MCP, retrieves the data, and answers the question. Instead of copying a Jira ticket description, you tell the AI "summarise the requirements for ticket PROJ-1234" and it reads the ticket directly.

The implications extend beyond convenience. When AI can access real-time data, it can perform tasks that were previously impossible: monitoring systems and alerting you to anomalies, keeping documentation updated as code changes, creating reports from live data rather than stale exports, and managing workflows that span multiple tools.

For organisations, MCP is the infrastructure layer that makes AI genuinely useful for enterprise work — where the most valuable data lives in specialised systems, not in the AI's training data and not in conveniently pasteable formats. It is the bridge between "AI that knows a lot of general things" and "AI that knows about my specific business."

How MCP works: the client-server model

MCP uses a client-server architecture — a well-established pattern in software that separates the requester (client) from the provider (server). Understanding this architecture helps you see how the pieces fit together.

The **MCP client** is built into the AI tool you use. Claude Desktop, Claude Code, and other MCP-compatible AI applications include an MCP client that knows how to discover and communicate with MCP servers. You do not need to build or install the client — it comes with your AI tool.

The **MCP server** is a small programme that connects to a specific external tool or data source and translates its capabilities into MCP's standard format. There are MCP servers for databases (PostgreSQL, MySQL, SQLite), file systems, development tools (GitHub, GitLab), project management platforms (Jira, Linear), knowledge bases (Notion, Confluence), communication tools (Slack), and many more. Each server exposes the capabilities of its connected tool — what data it can provide, what actions it can perform — in a standardised way.

The communication flow works like this. When your AI assistant needs information from an external source, the MCP client sends a standardised request to the relevant MCP server. The server translates this request into the format the external tool understands, retrieves the data or performs the action, and sends the result back to the client in a standardised format. The AI model receives the data and incorporates it into its response.

MCP servers expose three types of capabilities. **Resources** are data the AI can read — files, database records, documents, configurations. **Tools** are actions the AI can perform — creating tickets, sending messages, updating records. **Prompts** are predefined instruction templates that guide how the AI uses the server's capabilities. Together, these three capability types cover the full range of interactions an AI might need with an external system.

The protocol handles the complexities that make integration difficult: authentication (proving the AI has permission to access the data), error handling (dealing gracefully when the external tool is unavailable), and rate limiting (preventing the AI from overwhelming external services with requests). These details are managed by the protocol so that neither the AI model nor the end user needs to worry about them.

Available MCP servers: what you can connect today

The MCP server ecosystem has grown rapidly since the protocol's launch. Servers are available across several categories, and new ones are released regularly.

**Development tools.** GitHub and GitLab servers let AI access repositories, pull requests, issues, and code reviews. The filesystem server provides access to local files and directories. Database servers (PostgreSQL, MySQL, SQLite) let AI query data directly. Docker servers enable container management. These are among the most mature MCP servers and are particularly valuable for development workflows with Claude Code.

**Knowledge and documentation.** Notion, Confluence, and Google Drive servers let AI access your team's documentation and knowledge bases. This is powerful for research tasks — instead of pasting relevant documents into a conversation, the AI can search your knowledge base and pull the most relevant information automatically.

**Project management.** Jira and Linear servers give AI access to project boards, tickets, and sprint data. This enables workflows like "summarise the current sprint's progress" or "create a ticket for this bug I just found" without leaving the AI conversation.

**Communication.** Slack servers let AI read and post messages, search conversation history, and interact with channels. Email servers provide similar capabilities for email workflows. These enable the AI to operate as a participant in your team's communication flows.

**Data and analytics.** Database servers for major platforms let AI query your business data directly. Combined with the AI's analytical capabilities, this turns natural language questions into real-time business intelligence: "What is our customer retention rate by cohort for the last six quarters?" becomes a query the AI can answer from your actual data.

**Custom servers.** The MCP specification is open, and building a custom MCP server for your internal tools is straightforward for any development team familiar with basic API development. This means that even proprietary internal tools can be connected to AI through MCP. For a hands-on guide to setting up MCP servers with Claude Code, see the [MCP curriculum module](/coding/claude-code-mcp). The [AI Glossary](/glossary) entry on [Model Context Protocol](/glossary/model-context-protocol) provides additional technical detail.

Setting up MCP: a practical overview

Setting up MCP involves three steps: choosing your MCP-compatible AI client, installing the MCP servers for the tools you want to connect, and configuring the connections. The process varies by client and server, but the general pattern is consistent.

For **Claude Desktop** (the easiest starting point), MCP is configured through the application's settings file. You add server definitions that specify the server name, how to run it, and any environment variables needed for authentication. Claude Desktop then discovers the server's capabilities automatically and makes them available in your conversations. When you start a conversation, Claude sees which MCP servers are connected and can use them as needed.

For **Claude Code** (the development-focused tool), MCP servers are configured in the project settings or user settings. This allows project-specific MCP configurations — a web development project might connect to a PostgreSQL database and GitHub, while a data analysis project might connect to BigQuery and Google Sheets. Claude Code reads the MCP configuration alongside the CLAUDE.md file, giving it both project conventions and access to project-relevant external tools.

Most MCP servers are installed via npm (for JavaScript/TypeScript servers) or pip (for Python servers). The installation is a single command, and the server runs locally on your machine. This is an important architectural detail — MCP servers run on your machine, not on a remote service. Your data flows from the external tool through the local MCP server to the AI client. This local execution model gives you control over what data the AI can access.

Authentication varies by server. Database servers need connection credentials. GitHub servers need personal access tokens. Slack servers need bot tokens. Each server's documentation specifies the required credentials and how to provide them (typically as environment variables in the MCP configuration).

The learning curve is modest for anyone comfortable with command-line tools and configuration files. For non-technical users, MCP setup is currently best handled by an IT colleague or developer — the process is not complex, but it involves terminal commands and config file editing that can be unfamiliar.

Security considerations and best practices

MCP creates powerful connections between AI models and your tools and data. With that power comes responsibility — security must be considered carefully before connecting any sensitive system.

**The principle of least privilege applies.** When configuring MCP servers, grant the minimum access necessary. If the AI only needs to read database data, do not give the MCP server write permissions. If it only needs to access one GitHub repository, do not give it access to your entire organisation. Most MCP servers support fine-grained permission configuration, and you should use it.

**Authentication credentials must be protected.** MCP server configurations contain API keys, database credentials, and access tokens. Store these in environment variables or secure credential stores, never in plain text configuration files that might be committed to version control. Treat MCP credentials with the same care you treat any other system credentials.

**Review what data flows through MCP.** When an AI model queries a database through MCP, the query results are sent to the AI model for processing. Depending on your AI tool's data handling policies, this data may be processed on remote servers. Understand your AI provider's data handling commitments (Claude's enterprise plans include strong data privacy guarantees) and ensure they are compatible with the sensitivity of the data you are connecting.

**Audit MCP server code.** MCP servers run on your machine and have access to the credentials you provide. Use official, well-maintained MCP servers from trusted sources. For open-source servers, review the code before installing. For custom servers, follow your organisation's standard code review and security practices.

**Monitor usage.** Keep track of which MCP servers are connected, who has access, and what data is being accessed. Periodic reviews of MCP configurations ensure that connections remain appropriate as team members change, projects evolve, and security requirements shift.

**Start small.** Connect one or two low-sensitivity tools first (a test database, a public repository, a non-confidential knowledge base). Build confidence with MCP's behaviour and security model before connecting more sensitive systems. This cautious, incremental approach is consistent with the security-conscious AI adoption practices taught in the [Advanced curriculum](/school/advanced).

The future of MCP and what it means for AI

MCP is still in its early stages, but its trajectory points toward a fundamental shift in how AI integrates with business operations. Understanding where it is heading helps you plan for what is coming.

The near-term future (next 6-12 months) will see MCP server coverage expand dramatically. Today, servers exist for dozens of popular tools. Within the next year, expect coverage for hundreds — including major enterprise platforms like Salesforce, HubSpot, SAP, ServiceNow, and industry-specific tools across healthcare, finance, legal, and manufacturing. As coverage expands, the friction of connecting AI to your specific tool stack decreases.

The medium-term future (1-2 years) will see MCP become invisible infrastructure. Today, setting up MCP requires technical configuration. As AI tools mature, MCP connections will be configured through graphical interfaces, managed by IT departments through standard deployment tools, and offered as pre-configured packages by tool vendors. The protocol will fade into the background — users will not think about MCP any more than they think about HTTP when they browse the web.

The long-term implication is the most significant: MCP enables AI agents. An AI agent is an AI system that can take autonomous action across multiple tools and data sources to accomplish complex goals. MCP provides the infrastructure that makes this possible — the standardised connections through which an agent can read data, make decisions, and execute actions across your entire tool ecosystem.

Imagine telling an AI: "Prepare for tomorrow's board meeting." With MCP connections to your project management tool, financial data, CRM, and presentation software, the AI could pull the latest project statuses, generate financial summaries from live data, identify key wins and risks from CRM data, and draft presentation slides — all through standardised MCP connections. This is not science fiction; it is the logical endpoint of the infrastructure being built today.

For professionals and organisations, the practical takeaway is straightforward: begin building familiarity with MCP now. The professionals who understand how AI connects to business systems will be the ones who design, manage, and lead the AI-integrated workflows of the near future. The [Advanced](/school/advanced) and [Expert](/school/expert) levels of the Enigmatica curriculum cover AI agents, integration architecture, and the strategic implications of technologies like MCP.

Related Terms