We Eliminated $2,400/Year in SaaS Costs Using Claude MCP + Skills in One Afternoon
We replaced a $30-100/month content automation tool with Claude's Model Context Protocol (MCP) and Skills—in just 90 minutes. This blog walks through exactly how we built a system that fetches YouTube transcripts, researches competitive content, generates SEO-optimized articles, and publishes to our CMS—all without writing a single line of code. CTOs and digital leaders will see the practical steps, real constraints, and CFO-friendly business case for eliminating $2,400/year in SaaS costs while gaining full customization.
$2,400 per year. That's what we were about to spend on a content automation tool.
Then I spent 90 minutes testing Claude's Model Context Protocol (MCP) with Skills. Now we're not spending anything.
Here's what happened—and why your CFO will care.
The $100/Month Problem Every CTO Recognizes
We needed a system that could take YouTube videos, fetch their transcripts, research similar content, generate SEO-optimized blog posts, and publish them to our CMS.
Standard SaaS solution? $30-100 per month depending on usage. That's $360-1,200 annually for a single workflow.
But we're a digital transformation studio. We've built dozens of applications for clients. So I knew the real cost: $25,000-50,000 to develop this internally, plus 3-6 months of engineering time.
Both options failed my first rule of building: Don't be dumb.
What We Actually Built in 90 Minutes
I spent one afternoon setting up what I'm calling "Claude MCP + Skills"—a framework that lets Claude interact directly with external systems without building a traditional application.
Here's what it does:
Total development time: 90 minutes.
Total cost: $0 (uses existing Claude Desktop subscription).
Lines of code written: 0.
That's not a typo.
How MCP + Skills Actually Works
Model Context Protocol (MCP) is Anthropic's open standard that lets AI agents like Claude connect to external systems and APIs. Think of it as giving Claude hands to reach into your existing tools.
Skills are reusable instruction sets you create once and Claude follows every time. They're like Standard Operating Procedures, but for AI.
Here's the architecture we built:
MCP Servers We're Using:
The Skill We Created:
A 900-line instruction document that tells Claude exactly how to:
The skill reads our company context, brand guidelines, author info, and service pages from our Claude Project. Every time we give it a YouTube URL or article link, it executes all 13 steps automatically.
Why This Beats Building a Traditional App
I'm the founder of a digital studio. We build applications for a living. So when I say we're not building this as a traditional app, that's significant.
Here's what happened:
We started developing this as a Next.js application. API integrations, authentication, UI components, deployment pipeline—the full stack. We got 80% of the way there.
Then I realized I was violating my own rule: Find the fastest, most efficient way to get what you need delivered.
The traditional app had three fatal flaws:
1. It was static. Once we wrote the code, changing the logic meant rewriting algorithms. With Skills, we just edit text and the behavior changes immediately.
2. It couldn't adapt. If we wanted to add a new content source or change our writing style, we'd need a development sprint. With Skills, we update one document.
3. We couldn't talk to it. Once the app started processing, we couldn't intervene. With Claude, we can pause, give feedback, and course-correct mid-execution.
That last point is crucial. When Claude shows us five title options, we can say "I like #3, but when you write the article, emphasize the ROI angle more." The app couldn't do that.
The First Live Test: What Actually Happened
This video documents our first real test of the system. I'd written the Skill but never actually run it end-to-end.
I gave Claude a YouTube URL and said: "Create SEO-friendly blog post ready to publish on Webflow based on this YouTube video."
Here's what happened in the next 13 minutes:
Minute 1: Claude read the 900-line Skill documentation and understood all 13 steps.
Minute 2: It fetched the full YouTube transcript. (This alone took me days to implement correctly in the app version due to API limitations and stupid mistakes.)
Minute 3-5: It called Serper MCP multiple times to research similar articles, analyzing their structure and identifying content gaps.
Minute 6-7: It performed keyword analysis and generated five title alternatives.
Pause: I reviewed the titles, selected one, and gave additional context about the angle I wanted.
Minute 8-12: It wrote a 2,200-word blog post following our brand guidelines, including proper citations, internal links, and SEO formatting.
Minute 13: It published to Webflow as a draft.
First try. No errors. No debugging.
Compare that to the app version: I spent weeks fighting with markdown formatting issues, transcript API failures, and content generation inconsistencies.
The Business Case Your CFO Needs to See
Let's talk about the actual costs we're avoiding.
Traditional SaaS Option:
Build It Internally Option:
MCP + Skills Option:
The math isn't subtle. We're saving $2,400 annually compared to a mid-tier SaaS solution, and we're avoiding $25,000-50,000 in development costs.
But here's what your CFO really cares about: We can now apply this pattern to dozens of other workflows.
Email triage? MCP + Skills.
Lead qualification? MCP + Skills.
Report generation? MCP + Skills.
Each one saves $2,000-5,000 annually. That's $20,000-50,000 in operational savings without hiring additional staff.
What This Means for Your Development Roadmap
I've been in 13 years of board meetings where executives debate build vs. buy decisions. The conversation usually goes:
"We need this capability."
"Should we buy a SaaS tool or build it ourselves?"
"SaaS is faster but inflexible."
"Building is expensive but gives us control."
"Let's table this for next quarter."
MCP + Skills introduces a third option: Prompt it into existence.
Here's when it makes sense:
Good Fit for MCP + Skills:
Not a Good Fit:
The sweet spot? Those internal tools that have been sitting in your backlog for 6 months because they're important but not urgent enough to pull engineers off core product work.
How to Actually Implement This
You don't need to be technical to set this up. But you do need to think clearly about what you're trying to automate.
Here's our process:
Step 1: Install Claude Desktop
Download it from Anthropic's website. This is where MCP servers run.
Step 2: Configure Your MCP Servers
These are the connections to external services. You'll edit a JSON config file to add servers you need. Anthropic's MCP documentation has the full list and setup instructions.
For our blog automation, we're using:
Step 3: Create a Claude Project
Projects let you add context documents that Claude references. We added:
Step 4: Write Your Skill
This is the instruction document. Start with a clear workflow:
Write it in plain English. Be specific. Include examples of what good looks like.
Our Skill is 900 lines, but we started with 100 lines and refined it over three iterations.
Step 5: Test with a Real Example
Don't try to automate everything at once. Pick one workflow. Give Claude a real input. See what breaks.
In our first test, Claude tried to publish before getting title approval. We added one line to the Skill: "CRITICAL: STOP AND SHOW TITLES TO USER." Problem solved.
Step 6: Refine Based on Results
The beauty of Skills is you're just editing text. No code changes, no deployments, no testing environments.
We're now on version 3 of our Skill. Each iteration took 15-20 minutes to update.
The Limitations Nobody Tells You About
This isn't magic. There are real constraints.
You're limited by Claude's context window. Our Skill is 900 lines plus context documents. We're approaching the limits of what Claude can hold in memory at once.
MCP servers aren't always reliable. Third-party servers can have downtime or rate limits. We've had Serper MCP fail during searches and had to retry.
You need some technical literacy. Setting up MCP servers means editing JSON files and understanding basic config. Not coding, but not zero technical overhead.
It's not appropriate for mission-critical systems. We're using this for content creation, not customer-facing transactions. If something fails, we retry. That's fine for blog posts but not for payment processing.
Security matters. You're giving Claude access to external systems. Make sure you're using API keys with appropriate permissions. Don't connect it to systems with sensitive customer data unless you understand the implications.
These aren't dealbreakers. They're constraints you need to understand before committing.
What Happens When Everyone Can Do This?
Here's what keeps SaaS executives up at night: If companies can prompt their own automation tools into existence, why would they pay for generic solutions?
Klarna announced they're eliminating 1,200 SaaS tools and building their own with AI. That's not a small company making a bold bet. That's a public company with fiduciary responsibility making a calculated decision.
The trend is clear: Software-as-a-Service is becoming Software-as-a-Prompt.
Traditional SaaS companies are responding in three ways:
1. Embedding AI customization. Salesforce has Einstein GPT. Microsoft has Copilot in Power Platform. They're trying to make their platforms as flexible as custom-built solutions.
2. Doubling down on enterprise features. Security, compliance, uptime guarantees, support—things AI-generated tools can't easily replicate.
3. Acquisitions. Big players are buying AI automation startups to integrate the technology before it disrupts them.
For your business, this creates opportunity. You can now build internal tools that would've cost $50,000-100,000 with a few hours of setup time.
The companies that figure this out first will have a significant operational cost advantage.
What We're Doing Next
This blog automation was our proof of concept. Now we're expanding to other workflows.
Next up:
Each of these would typically cost $30-100/month as SaaS tools or require weeks of development time.
With MCP + Skills, we're estimating 2-4 hours setup time per workflow.
That's the real transformation. Not replacing one SaaS tool. Replacing dozens of them.
The Decision You're Actually Making
If you're a CTO, CPO, or digital leader, you're not deciding whether to try MCP + Skills. You're deciding whether to wait for your competitors to figure this out first.
Here's my recommendation: Pick one workflow that's been sitting in your backlog for 6 months. Something important but not urgent. Something that would cost $50-100/month as SaaS.
Spend 90 minutes this week setting it up with MCP + Skills.
If it works, you've saved $600-1,200 annually and proven a new capability to your CEO. If it doesn't work, you've invested 90 minutes learning why.
That's a bet worth making.
Need help getting started? Our 90-day transformation program includes automation strategy and implementation. We help CTOs identify which workflows are best candidates for this approach and set up the infrastructure to support it.
Or start with our free MVP program to test the concept with a single workflow before committing to full implementation.
Either way, the question isn't "Should we do this?" It's "How fast can we move?"
Your CFO is waiting for the savings. Your CEO wants to see results this quarter. And your competitors are already experimenting with this.
90 minutes. That's all it took us.
How much longer are you going to wait?
About the Author
Behrad Mirafshar is Founder & CEO of Bonanza Studios, where he turns ideas into functional MVPs in 4-12 weeks. With 13 years in Berlin's startup scene, he was part of the founding teams at Grover (unicorn) and Kenjo (top DACH HR platform). CEOs bring him in for projects their teams can't or won't touch—because he builds products, not PowerPoints.
Want Us To Build You An MVP For Free?
