November 18, 2026
Most MCP Tutorials Are Already Wrong. Here’s What Works Now.
You found an MCP tutorial. It looked clean, it had good comments, and the author clearly knew their stuff. You opened a terminal, installed…

By Jbird
6 min read
You found an MCP tutorial. It looked clean, it had good comments, and the author clearly knew their stuff. You opened a terminal, installed the package, pasted the first block of code, and hit run.
ModuleNotFoundError: No module named 'mcp.server.fastmcp'ModuleNotFoundError: No module named 'mcp.server.fastmcp'That's it. Line one. You haven't written a single tool yet and the whole thing is already dead.
If you're anything like me, the next hour goes like this. You check your Python version. You reinstall the package. You spin up a fresh virtual environment, because maybe it's a path thing. You search the error and find three forum threads with three different answers, one of which tells you to downgrade something you've never heard of. By the end you're wondering whether this Model Context Protocol stuff is just not for you.
It's not you. The MCP tutorial you copied is out of date, and right now most of them are in the same boat. I ran into this exact wall while building the servers for my own course, which is the reason I ended up checking every single line against the current SDK before I'd put my name on it. So here's what happened, what changed, and the small fix that gets most old code running again.
Why most MCP tutorials broke at the same time
Quick refresher if you're newer to this. MCP is the standard way to give an AI assistant like Claude access to tools and data. An MCP server is the small program you write to expose one of those tools. You're not building the AI. You're building the thing the AI uses.
MCP went from a niche announcement to an industry default fast, and in 2026 it grew up. On July 28 a new revision of the protocol shipped, and the official Python SDK went to version 2 right alongside it. A plain install now pulls version 2 by default, so anyone following an older guide gets the new SDK whether they meant to or not.
A major version bump is the SDK's way of saying "we're allowed to break things now," and they did. The core server class every beginner starts with was renamed, and the old import path wasn't kept around with a polite warning. It was removed. No alias, no shim. Old code just fails on the first line.
Now think about how much MCP content was written during the version 1 era. That was the stretch when everyone was getting curious, so that's when the blog posts, the videos, and the copy paste starter repos all got made. They're still out there, still ranking in search, still looking perfectly reasonable. Search results don't know a tutorial went stale. The one sitting at the top for "mcp tutorial" might be from a few months ago, which in MCP time is a long while.
That's not the authors' fault either. The material went stale faster than anyone could update it. That's what happens when a technology is young and popular at the same time. But it does mean the thing you're learning from might be quietly teaching you a version that no longer exists.
What changed in the MCP Python SDK v2
Here are the differences that bite beginners, in plain English.
The big one: FastMCP became MCPServer. For a long time nearly every Python MCP tutorial started with the same two lines, importing FastMCP from mcp.server.fastmcp and then creating a FastMCP object. In the current SDK that module is gone. The class is called MCPServer now, and it lives in mcp.server.
# Old (v1, fails on the current SDK)
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("Demo")
# Current (v2)
from mcp.server import MCPServer
mcp = MCPServer("Demo")# Old (v1, fails on the current SDK)
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("Demo")
# Current (v2)
from mcp.server import MCPServer
mcp = MCPServer("Demo")That's the error from the top of this article, solved. Two lines.
Your decorators mostly still work. This is the reassuring part nobody mentions. The lines that sit on top of your functions, things like @mcp.tool(), carry over with the same arguments. It's the class name and the import that break, not the tools you wrote. So if you've got an old server lying around, the fix is usually a lot smaller than that first error makes it look.
SSE is on its way out for remote servers. If a tutorial runs a remote server over SSE, it's teaching you a transport that's officially deprecated. Remote servers use Streamable HTTP now. Old SSE setups still limp along for the moment, but don't build anything new on them.
The protocol is stateless. Older guides spend time on sessions and connection setup. In the current protocol, every request carries what it needs on its own, so a request can land on any copy of your server without that copy remembering who you are. For a beginner that mostly means less to manage. For anyone running a server for real, it means scaling out gets a lot simpler.
The setup itself is quick. I work on Python 3.11 or newer and add the SDK with uv:
uv add "mcp[cli]"uv add "mcp[cli]"And here's the fastest gut check I know. Before you follow any MCP tutorial, look at the very first import. If it pulls FastMCP from mcp.server.fastmcp, it was written for version 1. You can still learn from it, you just know where it's going to break before it does.
If this kind of "here's what broke and why" breakdown is useful to you, hit subscribe on Medium so the next one lands in your inbox automatically. I'd rather send you one accurate post than ten fast ones.
What didn't change, and why that matters more
Here's the part that should calm you down.
Through every version bump, the core idea of MCP stayed exactly the same, because that idea is the whole point of the protocol. A server exposes three kinds of things. Tools are actions the AI can take, like looking up a ticket or checking a service. Resources are things it can read, like a file or a config. Prompts are reusable templates for common requests. That's the model. It held in version 1 and it holds in version 2.
The workflow held too. You build a server, then you test it in the MCP Inspector before you ever connect it to Claude. The Inspector lets you call your own tools by hand and see exactly what comes back, so when something's wrong you know whether it's your server or the AI side. That habit alone saves hours, and none of the churn touched it.
If you came up through IT like I did, this should feel familiar. It's networking fundamentals versus one vendor's command line. The vendor's commands change every couple of releases. How a packet gets from A to B doesn't. Learn the stable core once, and the renames turn into a ten minute fix instead of a lost evening.
That's the real case for learning MCP. The API moves. The skill doesn't.
Why now is the right time to learn it
I'll be straight with you. MCP is young and it moves, and it's going to keep moving. Anyone who tells you their guide is the final word on it is selling you something. Mine included, which is why I date stamp exactly what it's current to.
But that roughness is the opportunity. The people who got good at cloud while the consoles were changing every month are the ones who looked like obvious experts once things settled down. MCP is in that window right now. Most people bounced off a broken tutorial and moved on. If you push through the version change instead, you're already ahead of the crowd that quit at line one.
You also come out of it with something real to show. Not a certificate, an actual working server that an AI can use, something you can demo, put in a repo, and walk someone through. For folks trying to break into or move up in IT and security, that's a pretty solid thing to have in your back pocket.
Where to go from here
First, the easy one. Join the Jbird list for straight talk breakdowns like this one, written by someone who did the work and hit the errors first. I think you just have to hit follow on Medium and opt to receive a notification when I publish. I only am publishing once or a twice a week at most so I'm not going to flood your inbox.
Second, if you want the full path instead of a blog post's worth of it, I put it all into a course called Build Your First MCP Server. It's 90 pages and nine build along projects in both Python and TypeScript, taking you from "what's MCP?" to a working server Claude can use, all the way to publishing one on the official MCP registry so other people can install it. Every line of code is current to the 2026 v2 SDK. No AI background and no degree required, because you're building the tools the AI uses, not the AI. It's only 19.99 at the moment and you can check it our here to see if its up your alley (link below!)
Build Your First MCP Server!
Good luck out there, folks.