🤖 Modern AI Automation: MCP vs SKILL.md / agent.md (OpenCode)

A practical comparison for network & security automation engineers.

🎯 Introduction

As AI and Large Language Models (LLMs) evolve, the way we integrate them into real-world automation workflows is rapidly changing.

Two emerging patterns are gaining traction:

• Model Context Protocol (MCP) → Tool-driven, structured AI integration.
• SKILL.md / agent.md (OpenCode-style agents) → Prompt-driven, instruction-based AI behavior.

If you’re building AI-driven automation platforms (e.g., pyATS + Cisco + APIs + LLMs), choosing the right approach is critical for:

• Scalability
• Cost (token usage)
• Reliability
• Maintainability

In this blog, I’ll break down:

• What MCP and SKILL.md are
• Key differences
• Token usage implications
• When to use each
• Real-world architecture guidance

🧠 What is MCP (Model Context Protocol)?

Model Context Protocol (MCP) is a structured way to connect LLMs to real tools and systems.

Instead of prompting the model to “figure things out,” you expose tools like:

• Python functions
• APIs (Cisco FMC, SD-WAN, DNA-C)
• Automation frameworks (pyATS)

🔧 How MCP Works

1. User sends a natural language request.
2. LLM interprets intent.
3. MCP routes the request to a tool/function.
4. Tool executes (Python/API).
5. Result is returned to the user.

🧠 What is SKILL.md / agent.md (OpenCode Style)?

This is a prompt-engineering-based approach.

You define:

• Instructions
• Behavior
• Context
• Examples

…inside files like:

• SKILL.md
• agent.md

🔧 How It Works

1. LLM receives:
   • Prompt
   • Instructions (SKILL.md)
   • Context
2. LLM generates output directly.
3. No strict tool execution layer.

It’s more flexible, but less structured.

⚖️ MCP vs SKILL.md — Core Comparison

💰 Token Usage Comparison (CRITICAL)

This is where things get very important for real-world deployments.

🔹 SKILL.md / agent.md Token Usage

Each request includes:

• Full prompt
• Instructions
• Context
• Examples

Typical token consumption:

2,000 – 10,000 tokens per request

❗ Problems:

• Expensive
• Slower
• Repetitive context

🔹 MCP Token Usage

MCP minimizes token usage by:

• Sending only intent
• Offloading execution to tools

Typical token consumption:

200 – 1,000 tokens per request

✅ Benefits:

• Lower cost
• Faster execution
• Scalable

📊 Visualizing the Difference

SKILL.md Approach

User Request
   ↓
LLM (Prompt + Context + Instructions)
   ↓
Generated Output

Heavy token usage every time.

MCP Approach

User Request
   ↓
LLM (Intent only)
   ↓
Tool Call (Python/API)
   ↓
Execution Result

Minimal tokens, real execution.

🧪 Real-World Example

Scenario: “Check firewall health and fix critical issues.”

🔹 SKILL.md Approach

LLM must:

• Parse logs
• Generate logic
• Suggest fixes

❗ Risks:

• Hallucination
• Inconsistent output

🔹 MCP Approach

LLM:

• Calls run_pyats_health_check()
• Calls fix_firewall_policy()

Result:

• Accurate
• Repeatable
• Production-ready

⚠️ Key Limitations

❌ SKILL.md / agent.md

• High token cost
• Non-deterministic
• Hard to debug
• Risk of incorrect outputs

❌ MCP

• Requires backend development
• Needs tool integration
• Slightly more complex setup

🧠 When to Use What?

✅ Use MCP when:

• Building production systems
• Integrating with Cisco APIs (FMC, SD-WAN)
• Using pyATS / automation frameworks
• Cost optimization matters
• Accuracy is critical

✅ Use SKILL.md when:

• Rapid prototyping
• Knowledge-based tasks
• Documentation generation
• Simple workflows

🏗️ Recommended Architecture (Best Practice)

For modern AI automation systems, use a hybrid approach:

LLM (Chat Interface)
   ↓
MCP (Tool Orchestration Layer)
   ↓
Python / APIs / pyATS / Cisco Platforms

Optional:

• SKILL.md for the guidance layer
• MCP for execution

🚀 My Recommendation (From Real Experience)

Based on real-world automation (Cisco + AI + MCP):

MCP is the future for production AI systems.

Why?

• Lower token cost
• Higher reliability
• Better integration
• Scalable architecture

🔥 Key Takeaways

• SKILL.md = flexible but expensive and less reliable
• MCP = structured, scalable, and production-ready
• Token efficiency is a major deciding factor
• Hybrid approach = best of both worlds

🏁 Conclusion

As AI evolves, the shift is clear:

From prompt engineering → to structured AI systems.

If you’re building:

• Network automation platforms
• Security automation (FMC, FTD, SIEM)
• AI-driven operations

…then MCP should be your foundation.