Initial explorations with MCPs and Claude

Key Learnings About MCPs: When They Work (And When They Don’t)

I tried building a personal AI advisor using MCPs to access my 188 Notion journal entries. Here’s what I learned about their limitations:

The Problem: MCPs Don’t Scale for Large Datasets

What I needed: All 188 journal entries for comprehensive advice
What happened: 188+ individual API calls, eating tokens and going very slowly
Reality check: MCPs optimize for real-time queries, not bulk data retrieval

Where MCPs Struggle

❌ Bulk exports - Takes shortcuts, misses important data
❌ Comprehensive data analysis - Too expensive in tokens/time
❌ One-time data preparation - Not designed for this use case

Where MCPs Excel

✅ Real-time operations - Writing new entries back to Notion
✅ Targeted queries - “Get my latest reflection”
✅ Live integrations - Enhancing ongoing conversations

The Solution: Hybrid Architecture

Static export (direct Notion API) for historical context
+ Live MCP for recent entries and writing new content
= Best of both worlds

Key Insight

For AI advisors, front-loading context beats real-time retrieval. It’s more efficient to load your entire dataset once than make 50+ API calls every conversation.

Bottom Line

MCPs are great for enhancing conversations with live data, but they’re not built for the heavy lifting of comprehensive data preparation. Know which job you’re solving for.