What is MCP? A practical guide to the Model Context Protocol

What is the Model Context Protocol?

The Model Context Protocol (MCP) is an open standard that defines how AI agents communicate with external tools, data, and services. Think of it as a USB-C port for AI: any client that speaks MCP can plug into any server that speaks MCP, and the agent gains a new capability without writing custom integration code.

MCP was introduced by Anthropic in late 2024 and has since been adopted by Claude Desktop, Cursor, Continue, Windsurf, Cline, and a growing list of agent runtimes. The specification, SDKs, and reference servers are all open source.

At its core, MCP is a JSON-RPC 2.0 protocol that runs over stdio, HTTP, or Server-Sent Events. A client sends a request describing the tools it wants to call, and the server returns a structured response. Because the protocol is open, the same server can be used by multiple clients, and the same client can talk to multiple servers in one session.

How MCP works

An MCP session is a long-lived connection between a client and one or more servers. When the session starts, the client asks the server what it offers, and the server replies with a list of capabilities.

Servers can expose three kinds of capabilities:

• Tools — Functions the agent can call. A tool is described by a name, a description, and a JSON Schema for its inputs. Calling a tool returns a structured result that the model can use in the next step.
• Resources — Read-only data the agent can fetch. Resources are URI-addressed (for example, file:///repo/README.md) and can be text, JSON, or binary blobs.
• Prompts — Templated messages the user or the agent can insert into a conversation. Prompts are useful for common workflows like "summarize this issue" or "draft a release note".

The agent's language model decides which tool to call based on the user's request and the tool descriptions. The client forwards the call, the server executes, and the result becomes part of the model's context for the next turn.

What is an MCP server?

An MCP server is a small program that implements the Model Context Protocol. Most MCP servers wrap something the agent would otherwise have to talk to directly: a CLI, a database, a SaaS API, a local file system, or a remote service.

Common examples include the official GitHub MCP server, the Postgres MCP server, Playwright MCP for browser automation, and Filesystem MCP for local file access. There are also community servers for Notion, Slack, Linear, Stripe, Supabase, Sentry, and many more.

You can browse all of them in the MCP server directory. Each entry includes the install command, the source repository, the author, and the supported capabilities.

MCP vs plugins and tool APIs

MCP is often compared to plugin ecosystems like OpenAI's GPTs or the ChatGPT plugin format. The key difference is portability: an MCP server works with every MCP client, while a plugin is usually tied to a single product.

Other differences:

• MCP defines a structured tool-calling API, resources, and prompts in one protocol. Plugin formats usually only support function-style calls.
• MCP runs locally or in your own infrastructure. The server has access to the same resources the user has, with no data leaving the network unless explicitly configured.
• MCP is open. There is no central approval gate, and any developer can publish a server without going through a marketplace.

Which AI clients support MCP?

MCP works with the following AI clients today:

• Claude Desktop — The reference implementation. Configure MCP servers in claude_desktop_config.json.
• Cursor — Add MCP servers under Settings → Features → Model Context Protocol.
• Continue — Open-source IDE extension. Supports MCP servers through a YAML config.
• Windsurf — Built-in MCP support with one-click install for popular servers.
• Cline / OpenHands / Claude Code — Agent runtimes that consume MCP servers as their tool layer.

Because MCP is a standard, the list of compatible clients is growing. If a client does not appear here, check the project's documentation for "Model Context Protocol" support.

How to install an MCP server

Installing an MCP server is a three-step process. The exact steps depend on your client, but the pattern is the same.

1. Pick a server from the MCP server directory and read the install command on its detail page.
2. Register the command in your client's MCP configuration. For Claude Desktop, edit claude_desktop_config.json. For Cursor, use the settings UI.
3. Restart the client and verify the server's tools appear in the model context. You should see the tool names listed when the session starts.

Most MCP servers use one of four transport patterns: npx for TypeScript servers, uvx or pipx for Python servers, docker run for containerized servers, and wget for one-line installers.

MCP security best practices

MCP servers run on your machine or in your infrastructure, so they inherit the permissions of whatever process starts them. Treat them like any other piece of software you install from the internet.

• Review the source code before installing. MCP Haus links every entry to its public repository.
• Pin versions. Prefer @org/server@1.2.3 over @org/server@latest in production.
• Use the least-privilege credentials possible. For GitHub, prefer a fine-grained personal access token with read-only scopes. For databases, create a read-only user.
• Audit logs. Keep an eye on what your agent is calling, especially in shared environments.

Popular MCP use cases

Teams use MCP for a few recurring patterns. The most common are:

• Coding agents — Read repositories, run tests, and open pull requests from Claude Desktop, Cursor, or Continue.
• Data analysis — Let agents query a warehouse, summarize metrics, and generate reports on demand.
• Customer support — Connect agents to ticketing, CRM, and knowledge bases so they can answer with real context.
• Personal productivity — Wire agents to Notion, Linear, Slack, and Google Maps to coordinate the work day.
• Research and RAG — Index documents, papers, and web pages through search and knowledge servers.
• Platform engineering — Manage Kubernetes, Cloudflare, and Vercel resources through an agent interface.

Frequently asked questions