The MCP Myth

Do You Really Need a Unified Agentic Protocol?

Yan Wang, PhD

May 14, 2025

1
 “I've read the docs. Watched the demos. I get the
idea—but I still don't really know whether it applies
to what I'm building.”

"Worse, I've seen people use it too early, hit weird

bugs, and walk away thinking AI doesn't work. The
real issue wasn't the protocol — it was using it at
the wrong time."

2
Protocol Overload: Which One Should I Choose?

3
Live Demo: Tool Invocation with OpenAI

"Let's build an agent that queries a local database

using OpenAI's Function Calling."

4
 Agentic AI: Core Invocation Flow

Structured Tool
LLM Call Client Code Result
Decides to use a Specifies tool & Executes the tool Returned to LLM
tool arguments
e.g., Python, JavaScript e.g., JSON, text
e.g., GPT, Claude Schema-based
(e.g., JSON)

5
Wait—Same Tool, Different Model, Doesn't Work?

"I copied my tool from OpenAI to Claude... and it

broke."

6
 What Is the Model Context Protocol?
• A standardized interface for AI tools & resources.

• Abstracts LLM-specific implementation details.

• Uses JSON-RPC for communication.

• Promotes reusable, AI-friendly, interoperable tools.

• Establishes clear contracts for AI-tool interaction.

Core Goal:
Enable a rich ecosystem: develop tools once,
use with many AIs.

7
MCP Is Gaining Momentum

✅ Claude Desktop
✅ Cursor
✅ OpenAI
✅ Google
✅ Anthropic (creator)
✅ Community SDKs (FastMCP, etc.)

8
 Let's Build a Tool with FastMCP

Build a database query tool and expose it via

FastMCP. Then test it with both Claude Desktop and
Cursor.

9
 MCP Debugging: Not So Easy

We gave the MCP server an incorrect path for its

SQLite database… and got a really cryptic error
back.

10
You're Not Debugging Code. You're Debugging a
Protocol.

• Error messages o!en don't tell you what's wrong

• Client ↔ Server ↔ Model schema boundaries are hard to trace
• Abstract protocols reduce visibility during early experimentation

LLM Client MCP Server Tool

Decides Action Prepares Call
⚠ Processes Protocol Executes Task

11
Rebuild the Same Tool – Without MCP

Rebuild the same database tool using just a natural

language and a CLI

12
 MCP vs Prompt+CLI vs OpenAPI

Feature Function MCP Prompt+CLI OpenAPI

Calling

Cross-platform ❌ ✅ ⚠ ✅

Debug visibility ✅ ⚠ ✅ ⚠

Tool schema ✅ ✅ ❌ ✅
enforcement

Onboarding speed ✅ ⚠ ✅ ⚠

Ecosystem support Vendor-specific Rising Fragmented Mature (Open Web UI,

etc.)

13
Should You Adopt MCP Right Now?

• Are you targeting multiple LLM platforms?

• Do you need to standardize tool usage
across teams?
• Has your product reached a stable
deployment phase?

14
 MCP Helps You Scale. Not Start.

MCP is designed for robustness and reuse—not for speed of

iteration.

Concept
Exploration MVP Production
Prototyping

CLI + Prompt MCP

15
The Skill Isn't Coding MCP. It's Judging When to
Use It.

• Building agents is easy

• Knowing when to abstract is hard
• Builder's mindset is about knowing what matters now

"Premature optimization is the root of evil."

16
 A Builder Thinks in Three Dimensions

Why Hands-On Opinion

Understand trends Build end-to-end Filter out noise

Decode ecosystems Debug across tools Develop a point of view

Ask deeper questions Prototype rapidly Stay independent

17
The Full AI Builders Course: Build Judgment at
Scale

18
 Learn How to Learn. And How to Unlearn.

"This course not only teaches you practical skills,

but also trains your mind and establishes a
builder's mindset. In the era of AI, many people
ask: what should we do today? What skillsets
Max Xiao keep us relevant? I believe it's: learn how to
Sr. Principal Technical Program learn, and learn how to unlearn. These are not
Manager easily taught — but this course actually delivers
Oracle
them."

19
 AI Went From Tool to Habit.

“The greatest takeaway for me was the mindset

shi! — I used to think many tasks were beyond
AI's reach. Now, thinking about how to use AI has
become a habit. It's boosted my productivity
TT both at work and in life. This mindset shi! didn't
Applied Scientist just help me understand AI better — it changed
Microso! the way I think about solving problems.”

20
 I Built a Tool I Use Every Day.

“This course opened my imagination about what

I could use GenAI for. I built an application that I
now use daily — something I never thought I
could create so easily. It helped me think more
Christine deeply about what to focus on and what I can
Quant Researcher delegate to AI. It's not just about productivity —
JPMorgan & Chase it's about redefining how I work.”

21
 Basic Python. No Problem.

“With only basic Python knowledge, and working

in a non-so!ware company, this course still gave
me confidence and practical tools. The demos
were amazing and opened up possibilities I
Ken hadn't thought of. It made me realize I don't
Sr. Product Specialist have to wait until I'm 'ready' — I can start
ThermoFisher Scientific building now.”

22
 Is This Course Right for You?

✅ You've read the docs and still feel stuck

✅ You're building tools but unsure if you're on the right
track
✅ You want to learn how to think, not just copy code
✅ You want to talk about architecture, trade-offs, and
design—not just syntax

23
 Common Questions

Do I need to know Python? – No, but Will there be projects? – Yes, and
it helps you'll get feedback.

Is this for beginners or experts? – Is this course live or async? – Cohort

Both. It's about mindset. + lifetime self-paced access

24
 Thank You & Next Steps

Use coupon
code MCP for a
$149 discount
on the course.

"Misusing the right tool at the wrong time can cost more than not using it at all."

25