What is MCP? A plain-English guide to the Model Context Protocol

The short version

MCP (Model Context Protocol) is an open standard, originally published by Anthropic in late 2024, that lets AI models call external tools — databases, APIs, filesystems, browsers, design files, anything — through a single, predictable interface. Think of it as USB-C for AI agents. Before MCP, every model-tool integration was a one-off. With MCP, any compliant client can talk to any compliant server.

That's it. That's the whole pitch. Everything else is implementation detail.

Why MCP exists

Large language models are good at reasoning over context, but on their own they can't do anything. They can't read your database, can't see the file you have open, can't hit a real API. The first wave of fixing this was function calling: you, the developer, defined a JSON schema for each tool, the model emitted a function call, your code executed it, you fed the result back. It worked, but it didn't scale. Every model, every tool, every framework had its own glue layer.

MCP fixes that by separating concerns:

• The client (Claude Code, Cursor, Claude Desktop, etc.) handles the model and the user interface.
• The server exposes a set of capabilities (tools, resources, prompts) over a standard protocol.
• The user wires them together once. The client and server negotiate what's available on the fly.

Now any server can plug into any client. Once you've written a Postgres MCP server, every MCP-capable agent in the world can use it.

The three primitives

Every MCP server exposes some combination of three things:

1. Tools — functions the model can invoke. read_file, run_sql, send_message. These are the verbs.
2. Resources — read-only data the model can pull into context. A file's contents, a database schema, the current weather. These are the nouns.
3. Prompts — reusable templates the user can invoke. Think slash-commands surfaced to the model. /review-pr, /summarize-customer.

A simple server might just expose tools. A rich one (like the official Everything reference server) demonstrates all three.

Transports: stdio, SSE, and HTTP

MCP defines a JSON-RPC message format and three transports:

• stdio — the server is a subprocess of the client, talking over stdin/stdout. Simplest setup, no network, works great for local tools. Most servers in our directory default to this.
• SSE (Server-Sent Events) — the server runs as an HTTP service and streams events. Used for remote/hosted servers.
• Streamable HTTP — the newer transport that replaces plain SSE for production deployments. Bidirectional, stateless, scales.

What an MCP server actually is, in code

If you strip away the marketing, an MCP server is a small program that:

1. Listens for JSON-RPC messages on stdin (or HTTP).
2. Replies to initialize with the protocol version and what capabilities it offers.
3. Replies to list_tools with a JSON schema for each tool.
4. Replies to call_tool by actually doing the work and returning the result.

That's less than 100 lines of TypeScript or Python using the official SDKs. The reason there are now thousands of MCP servers is that the bar to write one is genuinely low.

The ecosystem in 2026

Today the MCP ecosystem has roughly four tiers:

1. Reference servers — maintained by Anthropic and the MCP project. Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, Time. These demonstrate the protocol and are battle-tested.
2. Official vendor servers — published by the vendor whose product they wrap. GitHub MCP, Stripe MCP, Notion MCP, Linear MCP, Atlassian MCP. These are usually the safest to use in production.
3. Community servers — written by anyone for everything else. Thousands of them. Some excellent, some abandoned. Check stars, last commit, and open issues.
4. Aggregators and gateways — meta-servers that route to many backends, or marketplaces that broker auth. Useful for fleets of agents but adds a layer.

Browse all 2,500+ servers in our directory →

How to start using MCP (today)

1. Install a client. Claude Code is the fastest path. Cursor, Cline, and Claude Desktop also work.
2. Pick a server. Start with something obviously useful — a database server, GitHub, or filesystem.
3. Register it. In Claude Code: claude mcp add <name> -- <install command>. Other clients have similar one-liners or config files.
4. Use it. Ask the agent something that requires the server. It'll discover the tools and call them.

See our Claude Code MCP guide for the full setup, including environment variables and credential handling.

What MCP isn't

• It isn't agent-to-agent. That's A2A (Google's emerging spec). MCP is model-to-tool.
• It isn't an auth standard. Servers handle their own credentials. The protocol doesn't prescribe how.
• It isn't magic. A bad tool description still produces a bad tool call. The protocol gives the model an interface — the design of that interface is on you.

FAQ

What does MCP stand for?

Model Context Protocol. Sometimes "Model Context Protocol Server" gets shortened to "MCP server" in casual conversation.

Is MCP only for Claude?

No. Anthropic published it, but it's an open spec. Cursor, Cline, Windsurf, Zed, Claude Desktop, and dozens of other clients support it.

Will MCP win as the standard?

It's the closest thing to a default right now. The ecosystem is growing fast. Whether it wins long-term depends on whether OpenAI and Google adopt it (Google has hinted) or push competing standards.

How do I find the right MCP server?

Use a directory. ClawHire indexes 2,500+ servers with install commands and category filters.