A few times a week, I catch myself whispering "holy crap"
A few times a week now, I catch myself whispering "holy crap" under my breath while AI does something I didn't think was possible.
I've watched AI fail a lot, but this week I had another "holy crap" moment while I watched Claude build a multi-platform software system using MCP connectors.
The client I built it for has a complex procurement process. It is burdensome even at small scale, but as the organization grows, it's not feasible to keep everything in check: following up on signatures; ensuring work orders don't expire; checking the invoice against the remaining budget; communicating to the vendors when a payment has been made. And on top of this, there needs to be an auditable paper trail for every step in the process.
As a fractional COO, part of my job is to help teams build processes and systems that allow them to scale sustainably, reducing bottlenecks, clarifying what often lives in one person's head, and, increasingly, building automations to reduce the cognitive load of the team trying to keep track of everything.
I mapped the whole process with them, step by step, and specified a web-based system that runs end to end. Here's what it does:
1️⃣ Helps decide what kind of procurement each request is
2️⃣ Routes it based on value, including how many competitive quotes are required
3️⃣ Captures the justification for the selected vendor
4️⃣ Generates the contract and the work order
5️⃣ Gives each vendor a portal to log in and see only their own files, submit invoices, and track status
6️⃣ Lets admin approve invoices and mark them paid.
At every step it logs the approvals and files PDFs into the relevant vendor folder so that there is an audit trail of the full process. It also gives admin a clear overview in one place: which invoices are outstanding, which work orders are expiring, and how much each vendor has spent against the budget.
I recently found an MCP connector between Claude and Lovable (one of my favorite tools to build front end web apps). Before this, I'd work with Claude to define a workflow. Claude would then write the prompts, and I'd paste them into Lovable one at a time, testing and adjusting and going back and forth until each piece worked. It always took a long time. With the connector, Claude builds directly in Lovable based on our product requirements document. I have the same setup with Supabase, the underlying database, and with n8n, which runs the triggers.
When I set all of this up last week, I was skeptical it would actually work. I've been through painful processes where Claude confidently builds something, only for me to find an endless list of errors. This time it worked beautifully. I'm still spending a lot of time on it: testing each step, asking for changes to make it more intuitive, getting the PDFs fully branded. But that's the QA work that goes into any software build. The difference now is that instead of working with developers, I'm working with Claude.
What MCPs are you using that are changing how you work?
This is one piece of my AI operating system.
I write about building it, in public, in my newsletter AI on Purpose.
Subscribe to AI on Purpose