Real talk from a technical founder building AI-powered businesses
Published September 26, 2025 • Based on Founder Reality Episode 22
Also available on: Apple Podcasts • Spotify • YouTube
The Infrastructure Shift Everyone Missed
While founders debate whether ChatGPT beats Claude or if Google Gemini will rule them all, something massive happened last week that makes all those arguments irrelevant.
GitHub and Microsoft just commoditized an entire industry. And if you're still building custom integrations for every AI tool you want to use, you're about to feel very stupid.
Here's what happened:
- GitHub launched the MCP Registry (free directory of AI connectors)
- OpenAI's responses API now supports remote MCP servers
- Microsoft shipped 10 MCP servers for developer workflows
- Entire startups offering paid MCP directories got nuked overnight
When Microsoft treats something as core infrastructure instead of an experiment, you should pay attention.
Let me explain what MCP servers actually are, why the scary name doesn't matter, and how this changes everything for founders building with AI.
What MCP Actually Means (No Technical Background Required)
MCP = Model Context Protocol
Think of it as a universal translator that helps any AI talk to your business tools.
Here's the simplest explanation: When you ask ChatGPT "What's the weather?" it can't actually check the weather because it's just a language model.
It needs to connect to a weather service. MCP is basically giving ChatGPT the phone number to call that weather service.
But here's the magic: Once you set up that connection, ANY AI can use it.
Not just ChatGPT. Claude, Gemini, Grok - they can all get weather data from the same connection.
You build the bridge once, and everyone can cross it.
The "server" part just means the connection runs in the background - like your WiFi router. You set it up once and it keeps working.
Real Business Examples (This Isn't Just for Tech Companies)
Every time there's AI infrastructure news, people assume it's only for SaaS companies or technical founders. Wrong.
Example 1: Dental Practice
Without MCP: You ask ChatGPT to check your client schedule. It can't access your booking system. You manually look it up.
With MCP: Your AI assistant has permission to check appointments, verify availability, send reminders - whatever you authorize it to access.
Example 2: E-Commerce Business
Without MCP: Every time a customer asks "Is this in stock?" you (or your VA) manually checks Shopify inventory.
With MCP: Your AI customer service agent checks your inventory system directly and gives real-time answers. No human needed.
Example 3: Consulting Business (My Setup)
Without MCP: When customers ask about billing issues, support manually checks Stripe, then responds.
With MCP: Our AI support immediately:
- Checks if they have an active subscription
- Sees if they cancelled or were overcharged
- Pulls relevant billing data
- Responds in seconds instead of hours
That's what people call "AI agents" - it's basically just an LLM plus MCP connectors.
The Week That Changed Everything
This wasn't just about new features. It was about infrastructure becoming commodity.
1. GitHub's MCP Registry
They launched an app store for AI connectors. Instead of every company building custom integrations with Notion, Stripe, Figma, and dozens of other tools, there's now a curated directory with big names already included.
Translation: You don't have to build everything from scratch anymore.
2. OpenAI's Remote Support
Previously, you had to keep your laptop running locally for your AI agents to use MCP servers.
Now: ChatGPT can just visit your website at something like yourbusiness.com/mcp/tools, authenticate, and start using your servers.
Translation: Your AI assistants work 24/7 without your computer running.
3. Microsoft Going All-In
When they're building MCP support directly into Windows and their enterprise tools, that's not an experiment anymore. That's infrastructure.
Translation: This is going to be everywhere. Learn it now or play catch-up later.
The Brutal Reality Check: Companies That Just Got Nuked
Here's the harsh truth: there were entire startups building paid MCP directories and marketplaces. Companies like Glamma, MCP.io, Composio - their whole business model was charging you to publish and connect MCP servers.
Then GitHub said: "Here's the same thing. It's free. It's where your code already lives."
Classic platform move. Take something emerging and bundle it to become the default.
This is why you never build where the platform can just roll in and give away your product for free. These companies now have to justify why they're charging money when Microsoft offers the same thing at no cost.
The Lesson for Founders
Don't build on top of platforms unless you have a defensible moat. Your moat isn't the technology - it's your:
- Proprietary data
- Network effects
- Brand trust
- Unique workflows
Microsoft can clone features. They can't clone your business relationships or proprietary processes.
How I Built Our SimpleDirect MCP System (Real Numbers)
I spent 5-6 hours over two days building a custom MCP integration that connects to Stripe.
What it does:
- Checks active subscriptions
- Verifies cancellation status
- Identifies overcharges
- Pulls any billing data we need
The ROI:
- Our entire team can use this connector
- Support response time dropped from hours to seconds
- Customer satisfaction increased (immediate answers)
- Zero ongoing maintenance (I built it once)
The strategic advantage:
Models are swappable, but connectors aren't.
We started with ChatGPT-5. We switched from Claude Sonnet 4 to Sonnet 4.5. When GPT-6 or GPT-7 comes out, I just switch the model and keep the same connectors.
I built this integration once, and it works with every future AI model. That's liberation from vendor lock-in.
The Security Reality Check (Don't Skip This Section)
Before you get too excited, let's talk about the risks, because MCP servers aren't the most secure systems on the planet.
Recent research shows they can be exploited for:
- Credential theft
- Remote code execution
- Data leaks
If you don't have proper authentication, anyone can visit your server and read everything in there.
My Security Protocol
Here's what I'm doing (you should too):
1. Read-Only Permissions First
AI agents CAN make edits if you allow them, which could be dangerous. Start with viewing only.
2. No Confidential Information Yet
I'm only connecting help center articles and stuff that's already public. Customer data requires more robust security.
3. Testing Everything
Even a 1% risk isn't worth taking with customer data. Test extensively before going live.
4. Don't Trust Connectors Blindly
Just because a connector is in the GitHub registry doesn't mean it's secure. Do your own research.
Security Checklist Before Going Live
- Implement proper authentication (OAuth 2.0 recommended)
- Set up rate limiting to prevent abuse
- Log all access attempts for monitoring
- Use environment variables for sensitive credentials
- Test with dummy data first
- Review permissions regularly (quarterly minimum)
- Have a kill switch ready if something goes wrong
Your Step-by-Step Action Plan (Do This Week)
Stop everything and try building one MCP connector. Here's exactly how:
Step 1: Start with Existing Tools (30 minutes)
Option A: Check GitHub's MCP Registry
- Go to github.com/mcp-registry
- Search for tools you already use (Notion, Stripe, Google Drive)
- Install pre-built connectors
Option B: Use AI's Built-In Connectors
- Claude already connects to Google Drive, Notion, etc.
- ChatGPT has Zapier integration
- Test what's available before building custom
Step 2: Set Up Read-Only Permissions (1 hour)
Don't give write access until you've thoroughly tested. Start with:
- Viewing customer data
- Reading appointment calendars
- Checking inventory levels
- Pulling support tickets
Step 3: Test Reusability (1 hour)
Connect two different AI agents to the same connector:
- Ask ChatGPT to check your Stripe data
- Ask Claude to check the same Stripe data
- Verify they both get consistent answers
This proves your connector is platform-agnostic.
Step 4: Document Everything (30 minutes)
Future you will thank present you. Write down:
- What connector you built
- What permissions you granted
- How to update/maintain it
- What problems it solves
- Security protocols in place
Step 5: Build Your First Custom Connector (2-4 hours)
Pick one business problem that requires manual checking:
- Customer billing questions
- Inventory status
- Appointment scheduling
- Support ticket status
Build a connector that solves it.
Expected outcome after this week: You'll be able to ask ChatGPT and Claude the same questions about your business data and get consistent answers from the same source. You'll be 2-5x faster than people using generic AI.
The Frameworks I Use to Decide What to Build
Not every process needs an MCP connector. Here's my decision matrix:
Build an MCP Connector If:
You or your team check it more than 5x daily
- Example: Customer subscription status
- ROI: Saves 10+ hours per week
Multiple people need access to the same data
- Example: Support team checking order status
- ROI: Eliminates information silos
The data changes frequently
- Example: Real-time inventory
- ROI: No more "let me check and get back to you"
Manual checking introduces errors
- Example: Billing calculations
- ROI: Reduces refunds and disputes
DON'T Build an MCP Connector If:
You check it less than once per week
- Not worth the setup time
- Better to just manually check when needed
The data is static or rarely changes
- Example: Your company's address
- Just tell the AI once in the prompt
High-security sensitive information
- Example: SSNs, passwords, financial credentials
- Wait until security protocols mature
You're the only user
- Unless it saves you 5+ hours weekly
- Otherwise, manual is faster
The ROI Formula
- Time to build: 2-6 hours for simple connector
- Break-even: When you save that time in the first month
- Compound value: Every month after = pure profit
Example: If building a connector takes 4 hours and saves 30 minutes daily, you break even in 8 days. After that, you save 180+ hours per year.
Advanced Use Cases (When You're Ready)
Once you've built your first connector, here are more sophisticated applications:
Multi-Tool Workflows
Connect multiple systems to create compound workflows:
- Pull customer data from Stripe
- Check support ticket history in Zendesk
- Review past orders in Shopify
- Generate comprehensive customer profile
One AI query = complete customer context.
AI-Powered Customer Support
Build a support agent that:
- Checks customer subscription status
- Verifies recent orders
- Reviews past support tickets
- Suggests solutions based on similar cases
This is what companies charge $500+/month for. You can build it in a weekend.
Internal Knowledge Management
Connect your:
- Company Wiki (Notion)
- Google Drive documents
- Slack conversations
- Email threads
Ask: "What did we decide about the pricing change in Q2?" and get instant answers with sources.
Automated Reporting
Build a connector that:
- Pulls revenue data from Stripe
- Fetches traffic from Google Analytics
- Gets email stats from ConvertKit
- Compiles weekly report automatically
No more manual spreadsheet updates.
The Infrastructure Shift You Can't Ignore
Most people are still debating which model is best. Meanwhile, GitHub just commoditized an entire industry and made AI infrastructure accessible to every founder.
Your business logic doesn't live in the models - those change every few months. It lives in the MCP connectors you build once and use forever.
Why This Matters for Long-Term Strategy
Traditional approach:
- Build custom integration for ChatGPT
- ChatGPT releases new model
- Rebuild integration
- Claude releases better model
- Switch to Claude, rebuild everything again
MCP approach:
- Build one connector
- Works with ChatGPT, Claude, Gemini, everything
- Model upgrades = just change the model, keep connectors
- New AI company launches = plug in your existing connectors
The companies that figure out MCP now will have AI systems that work across every platform. While everyone else is locked into specific vendors, you'll be free to switch between any AI provider while keeping all your custom integrations.
The Bottom Line
While everyone's fighting about ChatGPT vs Claude vs Gemini, the smart money is building connectors that work with everything.
Your competitive advantage isn't picking the right model. It's building the right infrastructure.
Models change every 6 months. Infrastructure compounds for years.
Build once. Use forever. Stay flexible.
Resources & Next Steps
This Week's Action Items
- Explore GitHub's MCP Registry for existing connectors
- Build one read-only connector for your most-checked data
- Test it with two different AI models
- Document your setup and security protocols
Discussion
Are you building with MCP servers yet? Drop a comment below - I'm actively responding to questions and sharing what I learn as I build these systems.
What connector would be most valuable for your business? Let me know and I might build a tutorial for it.
Are you building with MCP servers yet? Tweet me @TheGeorgePu - I'm sharing my experience building these systems as I learn.
New episodes of Founder Reality drop Monday/Wednesday/Friday at 9am EST. Real AI insights for builders, not startup theater. Full episodes and transcripts at founderreality.com
