One thing I have learned on the way to running a startup with a tiny team is that it is a constant battle against the clock. Every single day, there are a million things to do, and I am reminded constantly that my time, my resources, and my own know-how have very real limits.
One of the most frustrating things for me personally is dealing with repetitive tasks. You don't realize it at first, but a single weekly report can easily take me 1.5 hours to make. That's not even mentioning the mental load it takes just to get started. And if someone else on the team makes it, I often get a different quality than I expect.
Not only that, but because we wear so many hats, we rotate who does what quite often. If I have to do that same task again in 30 or 90 days, I might not exactly remember the steps. I end up spending even more time just figuring out where the data is supposed to come from in the first place.
We are a tiny team for a reason: a limited budget. A nightmare, right? Especially for anyone trying to make an impact in this world. Because of this, I want to utilize our available resources as efficiently as possible. I need lean business automation to help us achieve more.
In this series, I am going to share exactly what I am doing in my own company to drive better efficiency using AI.
My Target: Achieving More with Less
My goal isn't to automate our entire business and all of our processes overnight. Yours shouldn't be either. It’s about building a sustainable framework so you and your team can actually breathe and have the time to think about the strategic topics that help us grow.
Here is what I want to achieve from a business perspective:
- Reduce manual, repeating tasks: I want to cut down the number of repeat tasks we have to handle, which gives us the breathing space to grow our thoughts. For me, this means time to figure out strategic topics. For my team—and for yours—it means better designs and better requirements evaluation.
- Prevent knowledge leaks: When a process is either automated or properly documented, there is less chance that someone forgets how to do it a month later.
- Zero extra reporting effort: I want to standardize what can be standardized. I want us to rely on clean data rather than "gut feelings" or manual, typed-out update messages.
From a financial standpoint, I have a very specific set of constraints:
- Incur token costs only when it makes sense: We are already paying for tokens through our existing Claude and Gemini subscriptions. I only want to add more token costs if the automation truly adds extra value.
My Rule: Just-in-Time Automation
When you start researching the best automation tools, it’s easy to get carried away and try to automate everything. But for me, cost and time are major factors.
So, I made a strict rule for myself: First evaluate, start with Partial AI Automation including Humans, and only promote it to a fully automated workflow later.
You could try to fully automate things right from the start, but you wouldn’t really know how much value you are generating for the number of tokens you burn. Some workflows need humans in the loop anyway, and some just don't bring enough value to justify the setup.
Instead, my approach is to self-reflect on the tasks that are currently draining our effort. First, we figure out how much value we can actually generate by using AI. Then, we start doing those tasks with AI using our available subscription tokens. Finally, once we have a high level of trust and experience with the workflow, we move it into a fully automated solution.
Connecting the Dots: The Blueprint Behind This
While thinking about how to design a solution like this, a fair question to ask is: Why not just utilize all the free, open-source MCP servers out there?
For me, the answer comes down to three major roadblocks:
- Unmaintained Hobby Projects: Many open-source servers are just weekend projects. They don't get regular updates, and you can't rely on them for core business workflows.
- The "It Works on My Machine" Problem: Traditional MCP servers only run locally. If I set one up for Claude Desktop on my laptop, it won't work on the web version, and it won't work for my team. I would have to manually set up and maintain the servers across everyone's individual devices.
- Not Available on the Move: If I need to make a quick update or handle a task when I am away from my desk, I want to be able to do it seamlessly from my phone. Local servers completely lock you out of mobile access.
If those limitations don't bother you, then open-source MCP servers might be the right path forward. However, to fulfill my specific needs alongside our business goals, I ended up dividing our operational strategy into two distinct channels.
Two Paths: Thinking Work vs. Repeat Work
1. The Interactive Path (Me or the Team + Claude Web)
To begin with, almost everything starts here until we are confident the workflow actually works. For tasks that require deep analysis and strategic thinking, this path is permanent. We use system instructions and "skills" to keep our processes and institutional knowledge well-documented, ensuring the AI behaves exactly how we expect.
2. The Fully Automated Path (Events + n8n)
For completely repeating tasks—like that 1.5-hour weekly traffic report—I use n8n to catch events (obviously only after we've thoroughly tested and perfected the workflow in the interactive path). It calls AWS Bedrock to invoke the LLMs and handle the AI heavy lifting.
Why Bedrock? Honestly, because our existing infrastructure is already there, and as a lean startup, the last thing I want is to go through the administrative nightmare of approving and processing invoices from yet another software supplier.
Bringing It All Together
Ultimately, these conscious design choices allowed me to build a framework that fits our exact lifestyle:
- Utilize existing solutions as much as possible: I lean heavily on tools we already use, like n8n and Claude, rather than trying to build everything from scratch. The massive bonus here is that these platforms are improving at lightning speed—by building on top of them, our internal automation stack automatically gets smarter and more powerful over the next few years as they upgrade their own systems.
- Use MCP Express as a unified MCP server: This acts as our central bridge, saving me from the massive headache of managing and maintaining multiple separate servers.
- Skills and workflows for the win: By using built-in skills, we document and lock down our AI processes so the knowledge stays with the company, even if we forget the steps three months later.
This central infrastructure layer allows the AI to directly modify artifacts "in-place," increasing efficiency on tasks that would have otherwise taken a lot longer. Best of all? I am not chained to a local terminal on my laptop. I can access my tools, run a workflow, and check on my servers or work packages on the go, straight from my mobile while I’m away from my desk.
What’s Next? (And a Quick Favor)
There is a lot more coming in this series. Over the next few posts, I’ll be breaking down exactly how we improve our efficiency using AI and MCP.
In the meantime, here is how you can get started:
- Try it out yourself: Go ahead and register for MCP Express. We deliberately built a free tier that is more than enough for you to start experimenting and connecting on your own. You might want to grab this now if you intend to follow along with this series and apply these workflows to your own setup.
- Let us map your specific case: If you have a frustrating, manual process that is draining your time and you want to see if this lean framework can solve it, shoot an email to hello@mcp-express.com. I would love to look over your setup and potentially feature the solution in an upcoming post.
