Alt-tab is the devil
MCPs and the universal app
Model Context Protocol, or MCP, is a really good idea. It came out of Anthropic in late 2024 and makes it easy for AI Agents to connect to data and tools (an Agent is a neat little software trick to make LLMs plan and execute tasks autonomously).
Kind of surprisingly, other major AI shoppes also thought it was a good idea and were all on board within a couple of months.
People love MCP and we are excited to add support across our products - Sam Altman (CEO OpenAI)
To MCP or not to MCP, that is the question….MCP it is! - Sundar Pichai (CEO Google)
The application layer is collapsing into agents…Open protocols like MCP are key to enabling the agentic web - Satya Nadella (Microsoft CEO)
The problem they solve is that LLMs are little boxes of information with no access to the outside world. You can ask an LLM to summarize Shakespeare but not your bank transactions. It can help you write a grant application but can’t tell you who else received grants on your topic.
MCP makes it easy for developers to wrap existing databases and software into MCP “servers” that can then be plugged into any application that is an MCP “client”.
Anthropic uses the analogy of USB to explain this concept,
“Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect your devices to various peripherals and accessories, MCP provides a standardized way to connect AI models to different data sources and tools”.
Try to remember your car before you could connect your phone to it. You were stuck with whatever the car maker offered: radio, crappy navigation systems, and voice features that never worked. That’s AI without MCP.
The Great Cursor Unlock
If you’re a developer, you know Cursor. If not, it’s an AI-powered piece of software for writing code. The class of software is called an Integrated Development Environment (IDE) but that’s not important. What is important is that the company was started a couple of years ago by a few 20-somethings and is the fastest company ever to achieve $100M in annual recurring revenue.
A key reason for their astonishing success is that they beautifully integrated AI into the software engineer’s regular workflow. The engineer didn’t have to change, they just opened a familiar-looking app and had this massively powerful AI they could chat with. Before Cursor, engineers would alt-tab from their IDEs to a chatbot, ask it to write code, copy the code, alt-tab back into their IDE and paste it in. Then they would scrutinize it, test it, edit it, etc.
Cursor was an early adopter of MCP - I think they incorporated it within a month of Anthropic releasing the protocol. Now there are hundreds of MCPs to choose from, and some are actually useful.
Two MCPs that are incredible unlocks for my work are the GitHub MCP and AWS’s Log Analyzer MCP. The video below shows an example of how I use these together with Claude 4 Sonnet (one of Anthropic’s latest models).
It goes like this:
I created an Issue in my project’s GitHub complaining about burning cash in AWS and wondering if I could optimize a particular service.
From within Cursor, I tell the Agent to read the issue and look around the AWS logs for information.
The Agent first uses the GitHub MCP to read the issue.
It then uses the Log Analyzer MCP to find relevant logs in my AWS account.
It finds a log and tells me about its excitement before diving deeper.
After a few seconds, the Agent writes up its findings with a conclusion that there’s an easy optimization to save money.
What's not shown in the video is the next step where I asked the Agent to make the change and it rewrote the code for me.
Before Cursor, Agents, and MCPs, I would be alt-tabbing between different apps, logging in to each, clicking around to find what I needed, pulling it all together in my brain, and documenting my findings. And god forbid I forgot where I stored a password or how to navigate AWS logs (which happens all the time) - this job could take an hour. The video shows all this happening in under a minute.
What’s Next?
After experiencing the power of Agents with MCPs, it’s hard to look at all the apps open on my computer.
Not to mention all the tabs.
Taken together, these are just tools and information sources; the exact thing MCPs are built for. A future I’m looking forward to might come in phases:
Phase 0 - Alt-tabbing our way through life
This is where we are now.
Phase 1.a - Many popular apps add their own AI Agents
a la Cursor and Claude.
Phase 1.b - Agentic apps adopt MCP
Users can easily connect other tools and data sources without having to learn the specifics for each one.
Phase 2 - Most apps are both MCP clients and servers
Google Docs can summarize customer sentiment from my product database and Substack can turn my doc into a blog post; Photoshop can incorporate ideas from my favorite artists’ Instagram feeds and my team can see my designs come to life in Slack; Excel can make a budget spreadsheet based that email my CFO sent and fill in prices based on real-time negotiations with vendor agents; etc, etc.
Phase 3
Universal app. The un-app. Should it even be called an app?
The major app categories (create, consume, communicate, execute) converge and what’s left are the Agents, Data, and Tools. No more alt-tab. Whatever form(s) it takes, you won’t be paralyzed by zillions of icons and menu items. It will be made for everyone but personalized so each person can connect to their own stuff.
In this phase, the economics of software changes. Instead of seats and subscriptions, knowledge and its application become a consumable, a utility like energy (Sam Altman and others have talked about this). It’s not clear if we’ll see metered charges since different Agent/MCP providers will have different costs, but usage-based pricing seems plausible.
This brings up a bunch of other topics that may turn into more blog posts or a company :-)