MCP is about the right context, not just more context
Think in terms of source, freshness, and permissions.
Three core questions
Before connecting anything through MCP, ask:
- Source: where does this information come from?
- Freshness: is the model seeing live data, cached data, or a snapshot?
- Permissions: who can read, write, or only request approval?
Why this matters
Many failures are not reasoning failures. They come from bad context:
- stale data
- mixed versions of truth
- overly broad access
Start small
Expose the smallest resource set that can solve the current task.
Do not connect the whole internal system just because you can.