Video Notes: Understanding MCP (Beyond the Hype)
To get past the hype of MCP, I dove into a detailed interview with Tony Ramchandani, VP of Applied AI Engineering and author of an upcoming book on MCP.
Youtube Reference: https://www.youtube.com/watch?v=fuA8xHRnpXg
Here is what I learnt from the video:
Is MCP Only for Coders?
The short answer: No.
While developers build MCP integrations, making them work reliably at scale. Security teams need to manage access control, platform engineers need to host and orchestrate them, and QA teams need to evaluate how models interact with data.
Whether we write code daily or not, understanding MCP is quickly becoming part of the job.
The Building Analogy: How MCP Works
Think of an LLM (like Claude or ChatGPT) as a visitor walking into a secure office building:
- The Host/Client (The Receptionist): Checks the visitor in, verifies identity, and hands over a badge.
- The MCP Server (The Security Pass & Map): Gives the AI model clear, limited boundaries on where it can go and what it can touch.
Through this connection, MCP hands the AI three specific tools:
- Tools: Actions the model can execute (e.g., run a calculation, trigger a workflow).
- Resources: Read-only data it can inspect (e.g., documents, databases).
- Prompts: Pre-packaged templates and contextual instructions.
Instead of building 30 one-off integrations for 30 different tools, MCP gives you a single, standardized plug-and-play definition.
How MCP Compares: APIs, RAG, and Agents
- MCP vs. APIs: MCP doesn't replace APIs. In fact, most MCP servers wrap around existing internal APIs to give AI models a safe, standardized interface to call them.
- MCP vs. RAG: Retrieval-Augmented Generation (RAG) fetches documents to give an LLM extra context. MCP acts as the underlying protocol that connects those RAG tools to the model consistently.
- MCP vs. AI Agents: An agent is the "brain" making decisions and carrying out multi-step tasks. MCP is the "hands"—the standard plugin system the agent uses to interact with external tools.
Real-World Engineering: Securing Your MCP Setup
Connecting models to private enterprise data comes with real risks. If you're building or deploying MCP servers, keep these best practices in mind:
- Limit Active Tools: Don't expose dozens of tools to a model at once. Keeping active tools under 10–13 prevents context overload and keeps tool selection accurate.
- Use Hierarchical Tooling: If you have 50+ actions, group them logically. For instance, route
Read PDFandRead Text Filebehind a single parent tool likeRead Data. - Prioritize Security: Implement strict Role-Based Access Control (RBAC), enforce OAuth, mandate human-in-the-loop approvals for sensitive actions, and log every tool call for auditing.
How to Get Hands-On Experience
If you want to build a portfolio around MCP, avoid jumping straight into massive projects. Instead, follow a logical progression:
- Map the Architecture: Make sure you can clearly explain the Host, Client, and Server relationship on paper.
- Learn the SDKs: Pick up the official Python or TypeScript MCP SDKs.
- Build a Basic Server: Start small by creating a read-only server (like a basic calculator or local file reader).
- Layer on Security: Add audit logging, rate limits, and approval requirements.
- Ship a Domain Integration: Build a complete, enterprise-ready integration with full error tracing.
New Roles Emerging in AI
As MCP standardizes how AI connects to systems, new engineering specialties are taking off:
- AI Platform Engineers: Infrastructure experts managing MCP hosting and reliability.
- AgentOps / AI Reliability Engineers: The new DevOps for maintaining autonomous agent workflows.
- AI Evaluation & QA Engineers: Security and quality experts ensuring models execute tools safely without unexpected behaviors.
The Bigger Picture
MCP isn't the only framework gaining traction. Standards like Google's Agent-to-Agent (A2A) protocol, OpenAI’s
Agents.md, and the Goose framework are all pushing toward a more connected AI ecosystem.
