At my last doctor's appointment my GP showed me the results of an AI tool, summarising and automatically formatting our discussion into the clinical structure he needed. My exercise physiologist showed me something similar a few days later for recording assessments and creating exercise programs.
After each appointment I then spoke to a receptionist to pay and arrange a follow-up appointment. By contrast, this was a very repetitive and predictable process - the same for every patient. They checked available slots, swiped my private health insurance, and then processed my payment. For something so repetitive both receptionists seemed to make many mouse movements and clicks on their keyboards. These same manual tasks they've been doing for years and it seemed like nothing was optimised; perhaps it was even split across several systems. It was wild that health professionals were leveraging AI during appointments despite the privacy hurdles, while the admin tasks before and after remained manual.
All I could think was that their days were numbered.
The entire concept of needing a human to deal with this process and coordination of extremely common tasks (payments, bookings) could be very easily broken down into tiny little units of work that an LLM could invoke as needed. You could probably tap your phone on something on your way out and that'd be the end of it.
Sounds cool, but also far-fetched!
Well, last week I told you to focus on Domain Driven Design. This week I'm going to tell you what to build to pair with that knowledge to make this possible.
Anthropic quietly released something called MCP a while back. Most people missed it. Those who noticed thought it was just another protocol. Another standard in a sea of standards. They're wrong.
This Model Context Protocol is, in my mind, the foundation for a completely new software economy. One where your beautiful SaaS products become as relevant as DVD rental stores.
The great unbundling begins
Think about any SaaS product you use. Email marketing? Break it down. It's template storage, list management, sending capability, and analytics. Customer support? That's ticket routing, knowledge base search, response generation, and escalation logic.
We bundle these capabilities because humans need unified experiences. One dashboard. One login. One place to understand what's happening. Hilariously most businesses then subscribe to more than one. And then stitch things together, or manually battle with them every day.
AI doesn't need a unified experience to start with though. It can grab email sending from one tool, customer data from another, analytics from a third, and compose them into exactly what's needed. Right now today. For this specific situation. No subscription required.
LLMs have been able to invoke Functions or Tools since OpenAI shipped GPT 3.5. ChatGPT Plugins had a naive implementation and then GPT Assistants improved on them again. Anthropic's MCP standardised these tool calls and now the other AI APIs are adding support. This lets anyone build a tool but also, more importantly, use a tool. IDEs and chat apps allow them already and underlying frameworks like Cloudflare's AI workers and Vercel' AI SDK are making them trivial to build and consume in any product.
You might be thinking, sure, but will it take off?
WordPress plugins, but for AI
Remember when WordPress exploded? Suddenly anyone could add functionality to their site with plugins. Need a contact form? There's a plugin. SEO optimisation? Plugin. Social sharing? Plugin.
MCP servers are like WordPress plugins for AI. The difference? No user interface needed. They are small, focused, and do one thing well.
I've been experimenting with this myself. Just this week I prompted Claude with an idea for a blog series. It went away, conducted deep research, searched the web, came back and drafted three articles using Claude Artifacts. I chatted with Claude for a while afterwards, refining the articles via a Google Doc I had connected to my Project so Claude could use its in-built Google tools to read as I made updates. Once I was happy, I had Claude do a final spelling and grammar check and then we published the articles directly to my Webflow site via tools. Whilst I did use Google Docs to rewrite my document, I'm sure Artifact editing will arrive soon in Claude and then I would have never left Claude Desktop. No copy-pasting. No manual formatting.
The entire publishing workflow reduced to a conversation. In my mind, this is easier to see play out than WordPress. Once the MCP installation process is as easy as clicking a button across the main chat apps, it's going to take off like a rocket!
The economics are brutal
Let me show you what I mean:
Traditional SaaS model for a small business:
- Email marketing: $30/month
- CRM: $30/month
- Analytics: $30/month
- Customer support: $30/month
- Total: $120/month minimum
AI orchestrating tools:
- Send email: $0.01 per email
- Query customer data: $0.01 per lookup
- Generate analytics: $0.02 per report
- Route support ticket: $0.005 per ticket
- Total: Pay only for what you use
For a small business sending less than 1,000 emails and managing 500 customers the total cost is probably well under $30.
Even large SaaS companies in Silicon Valley won't need 10 to 20 different SaaS products in the very near future. You can finally put an end to all those recurring payments. Just wire up your business to direct tools and pay as you go.
And a small boutique might use email sending from one provider, SMS from another, customer data queries from a third, inventory checks from a fourth. All orchestrated seamlessly by AI that understands your business. Total cost? Easily less than one SaaS subscription.
OK, so where do I start?
Previously I said vibe coding should focus on creating atomic tools, not full applications. MCP is how you actually do that. It's the missing piece that makes the vision practical.
But which tools should you build? This is where last week's Domain Driven Design knowledge becomes crucial. When you understand the events and contexts in a business domain, the tools practically identify themselves.
Take my physio's clinic. After an event storming session, you'd discover events like "appointment scheduled", "treatment plan created", "insurance claim submitted", "payment processed". Each event needs specific capabilities. Insurance claim submission? That's a tool that validates provider numbers, formats claims to health fund specifications, and submits them. Payment processing? A tool that handles HICAPS integration for instant health fund claims.
Or consider work in retail. Events like "stock received", "price updated", "sale started", "return processed". Each bounded context reveals tool opportunities. The pricing context needs tools for competitor price checking, margin calculation, and promotional pricing rules. The inventory context needs stock level queries, reorder triggers, and supplier communications.
Even my old government work had clear tool needs. "ABN registered", "tax return lodged", "payment plan created". Thousands of micro-capabilities that currently live inside monolithic systems, waiting to be extracted.
I built a desktop app recently with a friend to explore this concept. Think MSN Messenger but each contact is an AI assistant with different tools enabled. One could search the web, another could create documents, and a third could generate images. Each assistant had unique superpowers based on the tools we'd given them. The whole thing took a few weeks to build, including creating tools to manage collections and items in an iCloud backed CMS.
I've also been using Windsurf (IDE) lately and hooking up tools to do more during my coding sessions. Simple things like XCode tools that let me build and open the Simulator without ever touching XCode. Each solving one specific friction point in my workflow.
The Australian business domain is ripe for tools. Every Australian business deals with GST calculations, ABN validations, and PAYG withholding. These aren't sexy problems, but they're universal in this market. "Build a tool that calculates Australian GST for digital products" could be vibe coded and launched in a few hours. Or take it a few steps further - 3 tools for the server: 1. ABN Validator, 2. ABN Search, 3. The GST calculation. Hyper-specific and relevant perhaps only to Aussies, but you can start to see how small a tool can be and still provide value. Value that people would pay to use too. And the code to build? No team. No sprints. No standups. Just vibe code it over a weekend and pay for the hosting if you can't find a free provider for the scale you need. The goal though - clear intent translated into focused capability. Quickly.
And the tools are embarrassingly simple.
Extract colours from a website. Validate a phone number. Recommend social media image dimensions. Each tool does exactly one thing. But they are powerful when combined with an LLM and a handful of other tools. Imagine an Australian ABN / GST MCP Server when you also add Stripe's MCP Server. Now you can process payments, using natural language.
Hang on, what about my SaaS?
If you're running a SaaS company, you have a choice. Wait for someone else to unbundle your features into micro-tools, or do it yourself.
Look at your product through a DDD lens. What are the actual events your customers care about? Which capabilities do they use daily versus once a month? That daily-use feature hiding in your settings menu? That's your first MCP tool.
Pick your best feature. The one customers actually use. Extract it. Make it bulletproof. Expose it via MCP. Price it at 1/100th of your current per-user cost.
Your customers will save money and you'll stay in business. Everyone wins except the old model.
The shift is already happening. While everyone debates AI replacing jobs, developers are quietly building the micro-tools that will power tomorrow's AI orchestrations. Every SaaS product is about to become a collection of API calls. Every workflow is about to become an AI conversation. Every monthly subscription is about to become usage-based pricing.
So if you're in the creator game, where's your MCP server? What capability are you exposing? What tool are you building? Stop your year-long side project, the bundling era is over. AI doesn't care about your dark and light theme toggle. Skip it. The tool economy has begun!