I remember the exact moment I realised I was working for my computer, rather than the other way around. It was late on a Tuesday. I had a bug in a Python script. I did what we all do. I opened the file. I highlighted the code. Ctrl+C. I Tabbed over to Claude. Ctrl+V.
Then I realised I forgot the context of the library it was using. Back to the editor. Open requirements.txt. Copy. Tab. Paste.
Then I realised it needed the error log. Open a terminal. Copy. Tab. Paste.
I was not a developer anymore. I was a glorified data pipe. I was manually ferrying bytes from one window to another because these two brilliant pieces of software… my editor and my AI… refused to speak to each other. They were like two geniuses sitting in the same room, ignoring each other, while I ran back and forth passing notes.
We have accepted this friction as the cost of doing business. We have accepted that “Context” is something we must manually curate, chop up, and feed into a text box.
That ends now.
The Invisible Wall
For the last two years, we have been obsessed with “More Parameters” and “Larger Context Windows”. We celebrated when models went from 8k tokens to 100k, then to a million. We thought the solution to the “dumb AI” problem was just to give it a bigger brain.
But a big brain is useless if it is locked in a padded room.
That is what ChatGPT, Claude, and Gemini have been. They are locked rooms. They know everything about the internet up to their training cutoff, but they know nothing about your hard drive. They do not know about the draft email you started yesterday. They do not know the schema of your local SQLite database.
To fix this, companies started building “Integrations”. You know the drill. You grant permission for a bot to read your GitHub. Then you grant another permission for it to read your Slack. Then another for Google Drive. It is a security nightmare and a development slog. Every integration is a new bespoke bridge that has to be built, maintained, and secured.
It is messy. It is exhausted.
Enter the Model Context Protocol
Anthropic recently released something that looks boring on the surface but is actually revolutionary. It is called the Model Context Protocol (MCP).
If you are not a systems engineer, that name sounds dry. It sounds like another acronym to ignore. Do not ignore it.
Think of MCP as USB-C for AI.
Before USB-C, we had a drawer full of cables. A cable for the printer. A cable for the monitor. A cable for the hard drive. If you bought a new device, you needed a new driver and a new cable.
MCP does for AI data what USB-C did for hardware. It creates a standard socket.
Here is the simple version.
You have a “Host” (like the Claude Desktop app or an IDE like Zed).
You have a “Server” (a tiny program that sits on your computer and knows how to talk to a specific thing, like your file system or a Postgres database).
You have a “Client” (the AI model).
They all agree to speak the same language. JSON-RPC. It is a simple text-based way for programs to say “Hello” and “Here is the data.”
Because they agree on the standard, you do not need to wait for Anthropic to build a “Google Drive Integration”. A developer, or you can, write a simple script (an MCP Server) that says “Hey, I know how to search Google Drive.”
You plug that script into Claude. Suddenly, Claude can search your Drive. You did not paste anything. You just asked.
Why This Feels Different
I installed the MCP server for my local file system last week. I opened the Claude Desktop app. I did not copy code.
I just typed, “Analyse the rust files in the src folder and tell me why the build is failing.”
The AI did not hallucinate. It did not ask me to paste the code. It “raised its hand” (a process called sampling) and asked the MCP server… “Can I please read the files in src?"
The server said yes. The AI read the files. It found the bug.
It felt like magic. But it was not magic. It was just… a connection.
For the first time, the AI felt like it was in the room with me. It was looking at the same screen. It had access to the same tools. I was no longer the courier. I was the director.
The Death of the Wrapper
This shift has massive implications for the industry.
For a long time, startups have been building “AI Wrappers”. These are apps that are basically just GPT-4 with a custom prompt and a custom connection to a database. “Chat with your PDF”. “Chat with your Excel sheet”. “Chat with your Notion”.
MCP kills that business model.
If I can just run a local MCP server that connects Claude directly to my PDFs, why do I need your monthly subscription? The “feature” of connecting to data is no longer a product. It is a protocol.
This returns power to the user. It allows us to build small, sharp tools that do one thing well.
You can write an MCP server that just checks the weather.
You can write one that just queries your smart home devices.
You can write one that checks the ticket prices on a specific airline website.
And once you write it, it works with any AI that supports the protocol. You write it once. You use it everywhere.
The Unix Philosophy Returns
This is what excites me most. It is a return to the Unix Philosophy.
In the Unix and Linux world, we have a saying… “Write programs that do one thing and do it well. Write programs to work together.”
The modern web forgot this. We built massive, walled gardens. We built “Super Apps” that tried to do everything.
MCP brings us back to modularity. It encourages us to build tiny “servers” that expose our data, without building a whole user interface around them. The AI is the user interface.
It Is Still Early
I must be honest. This is bleeding-edge technology.
Setting up MCP servers right now requires you to open a terminal. You have to edit configuration files. It is not “One-Click” yet.
But the momentum is undeniable. Companies like Block (Square) and Replit are already adopting it. The open-source community is churning out servers for everything from Git to Spotify.
We are witnessing the transition from “Chatbots” to “Agents”.
A chatbot talks. An agent acts.
But an agent cannot act if it is blind and helpless. MCP gives it eyes to see your files and hands to use your tools.
The Human Element
We often fear that AI will replace us. That is a valid fear if we view AI as an entity that lives in the cloud and sends us orders.
But when I use MCP, I feel differently. The AI has moved onto my machine. It is working within my environment, under my rules, as I look at my messy folders.
It makes the technology feel less alien. It feels less like an Oracle speaking from the heavens, and more like a junior developer sitting at the desk next to me.
We are moving away from the era of “Prompt Engineering”, where we have to trick the robot into understanding us. We are entering the era of “Context Engineering”. We simply give the robot the keys to the library and let it read.
I am ready to stop copy-pasting.
I have better things to do with my hands. Are you too?