Often I scroll through LinkedIn, it feels like walking through an AI‑themed amusement park.
There’s always someone proudly posting a glossy screenshot:

“I built my entire Power BI dashboard with MCP servers and AI!”
“Paginated reports? Done in minutes thanks to AI!”
“My ticketing system now has a chatbot that basically runs the company!”

It all looks so smooth, so elegant, so… suspiciously effortless.

And of course, I wanted that too.

In my head, the dream was clear: an AI apprentice sitting right next to me, integrated directly into SSMS and Power BI Desktop. MCP servers acting like a universal translator between my tools and my ideas. Agents automating repetitive tasks while I sip coffee and pretend to be strategic.

A beautiful dream.
A very distant dream.

The Missing Manual Nobody Talks About

The first real obstacle?
There is no structured walkthrough for any of this.

Sure, there are screenshots.
Beautiful ones.
Screenshots that whisper, “Look how easy this is.”

But behind the curtain?
Silence.
No reproducible steps. No architecture. No troubleshooting. Just vibes.

So I jumped in myself—motivated, curious, and armed with a healthy dose of skepticism.

Corporate Reality Check

My company only has Copilot through O365.
No GitHub Copilot.
No M365 Copilot.

And we’re strict about data governance:

• PII must never leave the company
• I don’t want my data used for training
• And paying 15–20 dollars a month for a cloud AI solution is hard to justify when you can’t yet measure the value

So the idea of a local AI setup became even more attractive.

First Taste of Success: Claude Desktop Free

Claude Desktop Free gave me my first “wow” moment.
Paginated reports?
Done quickly.
Surprisingly good results.
Combined with Beth Maloney’s rdl-mcp server, it felt like I had finally cracked the code.

But then reality tapped me on the shoulder:

• Claude Desktop isn’t local—it’s cloud‑based
• The output speed is… meditative
• And after one hour of joyful experimentation, I hit the 5‑hour prompting limit

So much for momentum.

The Big Experiment: Building a Local AI Stack With AI’s Help

I decided to go full inception:
Use AI to help me build a local AI solution.

Installing Ollama and running Phi‑3?
Easy. Encouraging. Almost relaxing.

But integrating MCP servers locally?
That’s where the universe said, “Let’s make this interesting.”

Copilot kept suggesting tools that sounded promising but behaved like experimental art projects.

Continuity (VS Code Plugin)

• MCP server not detected
• Model hallucinating like it had just woken up from a fever dream
• Zero real integration

Lobe Chat

This one deserves its own chapter.

I didn’t just “spend four hours.”
No.
I invested four hours the way people invest in questionable cryptocurrencies—full of hope, denial, and the faint suspicion that I should have known better.

• The “Windows version” is basically a fossil
• The Docker version worked after 1.5 hours of coaxing
• Ollama ran fine
• MCP still refused to cooperate
• I even built a local version with outdated Node.js 18
• It ran… but MCP was still “not available,” like a colleague who’s always “in a meeting”

Copilot: The Encouraging Partner in Crime

Throughout all this, Copilot was incredibly supportive.

“Martin, we’re just one step away.”
“This is a very important insight that moves us forward.”

Four hours later, we were still “one step away,” which made me wonder whether this step was actually the size of the Himalayas.

And I was deep in areas I normally avoid:

• Docker deployment
• Node.js development
• Webservice wrappers
• Tool integration

Copilot kept saying things like:

“No problem, we’ll just write a quick Node.js wrapper service.”

In practice:

• The webservice ran
• MCP still didn’t
• And querying a database through this “quick wrapper” was pure science fiction

The Slow, Slightly Painful Realization

After enough trial and error to qualify as a personality trait, the truth became clear:

• Cloud AI isn’t expensive just because of hardware
• Running a fully functional AI environment locally is not plug‑and‑play
• Ollama with a model is easy—but MCP servers and tool integration are a different sport
• And my modest Nvidia GPU?
  Let’s just say it’s not exactly eager to run GPT‑5.
  It doesn’t run slowly—it doesn’t run at all

Conclusion: The Honest, Slightly Funny State of Local AI

Local AI is exciting, empowering, and absolutely worth exploring.
But it’s not the magical, screenshot‑ready experience LinkedIn sometimes suggests.

You can run models locally.
You can experiment.
You can build cool prototypes.

But a fully integrated, MCP‑enabled, tool‑aware AI assistant that rivals cloud solutions?
That’s still a challenge—especially without deep experience in Docker, Node.js, and system integration.

And that’s okay.

Because every failed attempt teaches you something valuable.
And sometimes, it even teaches you to laugh at yourself a little.

If anything, this journey made me appreciate both worlds:

• The privacy and control of local AI
• The power and convenience of cloud AI

And the truth is:
Right now, you often need a bit of both.

And hey — if someone out there actually understands this stuff far better than I do, or if I somehow missed a beautifully simple guide hiding in plain sight, feel free to drop a comment right here. I’m always happy to learn… especially from people who didn’t end up waving the white flag like I did.

Disclaimer

This article was created based on my personal notes with support from Microsoft Copilot. While Copilot assisted in structuring and refining the content, all technical details have been carefully reviewed and developed by me.