Every AI application that wants to read files, query a database or call an API needs custom integration code — and every tool vendor must build separate integrations for every AI app. The Model Context Protocol (MCP), introduced by Anthropic as an open standard in November 2024, aims to solve that many-to-many problem. Its documentation describes it as a “USB-C port for AI applications”: build a connector once, use it with any compatible app.

The problem MCP solves

Before standards, connecting M AI applications to N tools required up to M × N custom integrations. With a shared protocol, each application implements MCP once, each tool exposes an MCP server once, and they all work together.

MCP was quickly adopted beyond Anthropic: other major AI companies, developer tools and code editors added support, and thousands of community servers appeared.

The architecture

MCP uses a client–server model:

  • Host: the AI application the user interacts with — a desktop assistant, an IDE, an agent framework.
  • Client: the component inside the host that maintains a connection to a server.
  • Server: a program that exposes capabilities from some system — a file system, GitHub, a database, a SaaS product.

Servers can run locally (communicating over standard input/output) or remotely (over HTTP), with authorization for remote access.

What servers expose

MCP servers offer three main kinds of capabilities:

  1. Tools — functions the model can call, such as “create issue”, “run query” or “send message”.
  2. Resources — data the application can read into context, such as files or records.
  3. Prompts — reusable templates for common tasks.

When connected, the host lists the server’s tools to the model. The model decides when to call one; the host executes it through the server — usually after asking the user for permission — and returns the result.

Trying MCP

  1. Pick a host that supports MCP, such as Claude’s desktop app, Claude Code or a compatible code editor.
  2. Choose a server — the official repository lists reference servers for file systems, Git, fetching web content and more, and many companies publish official servers for their products.
  3. Add it to the host’s configuration, typically a JSON entry with the command to start the server or the URL of a remote server.
  4. Ask the assistant to use it: “list the open issues in this repository” or “summarize the files in this folder”.

Building your own server

Official SDKs exist for several languages, including TypeScript and Python. A minimal server defines a few tools with names, descriptions and input schemas, and implements what each one does. Clear descriptions matter: the model reads them to decide when to use each tool.

Security considerations

MCP connects AI to real systems, so treat servers like software with access to your data:

  • Install servers only from trusted sources — a malicious server can read data or run commands.
  • Review permissions and prefer read-only access where possible.
  • Watch for prompt injection: data returned by tools (web pages, emails, issues) may contain instructions aimed at the model.
  • Keep approval prompts on for actions that modify data or send messages.
  • Protect credentials that servers use, and scope tokens narrowly.

Common mistakes to avoid

  • Installing many servers “just in case”. Each server adds tools the model must choose between and expands the attack surface. Enable only what you need.
  • Vague tool descriptions in your own servers. The model relies on descriptions to choose tools; ambiguous ones cause wrong or missed calls.
  • Auto-approving write actions. Keep human confirmation for anything that changes data, spends money or contacts people.
  • Using long-lived, broad tokens. Give servers credentials with the minimum scope and rotate them.

Key takeaways

  • MCP is an open protocol, introduced by Anthropic in November 2024, for connecting AI apps to tools and data.
  • Hosts contain clients that connect to servers exposing tools, resources and prompts.
  • One server works with every compatible AI application, replacing custom integrations.
  • Servers can run locally or remotely, and SDKs make building them straightforward.
  • Treat servers as privileged software: trusted sources, minimal permissions, human approval.

Frequently asked questions

Is MCP only for Anthropic’s Claude?

No. It is an open standard, and support has spread to many AI applications, development tools and agent frameworks from different companies.

Is MCP the same as function calling?

Function calling is how a model outputs a request to use a tool. MCP standardizes how applications discover and connect to tools, so the same tools work across apps.

Do I need to code to use MCP?

Not to use existing servers in supported apps; usually you edit a configuration file or click to add a connector. Building new servers requires programming.