When to Use MCP?
MCP = LLM lifeline when models hallucinate most.
Hallucination Triggers (Use MCP)
1. Knowledge Beyond Cutoff
❌ “2026 Q1 earnings?” → “Made up numbers”
✅ MCP: get_earnings(“2026‑Q1”) → Real API/DB
2. Ambiguous/Edge Prompts
❌ “Fix login bug” → Generic broken code
✅ MCP: validate_user_login(email, password) → Your auth logic
3. High Temperature/Creative
❌ Temp=1.0 → Creative but wrong test cases
✅ MCP: generate_test_cases(product_id) → Your test suite
MCP Shines Here
| Scenario | Without MCP | With MCP |
|---|---|---|
| Real‑time data | “Stock price?” → stale/guess | get_stock_price("AAPL") |
| Private data | “My Q4 sales?” → hallucinated | get_user_sales(user_id) |
| Complex logic | “Risk score?” → simple formula | calculate_risk_score(params) |
| Test automation | “Edge cases?” → invented | get_edge_cases(product) |
Code Patterns
Real‑time external data:
@mcp.tool()
def current_weather(city: str) -> dict:
return api.openmeteo(city) # Fresh data every call
User‑specific data:
@mcp.tool()
def user_orders(user_id: str) -> list:
return db.user_orders(user_id) # Private data
Business logic:
// Schema enforces valid inputs
{
"risk_score": {
"inputSchema": {
"properties": {
"age": {"type": "number", "minimum": 18},
"income": {"type": "number", "minimum": 0}
}
}
}
}
Don’t Use MCP When
- Static knowledge (Shakespeare quotes)
- Pure generation (creative writing)
- Simple math (2+2)
MCP = External truth + complex logic.
Call‑to‑Action
“Know WHEN to use MCP. Next: WHERE hallucinations hide worst.”
