Great article! I love MCPs and use them constantly, but you really have to be careful of how many you use at a time.
It happened to me more than once to have Sonnet call 4 servers one ofter the other as it was not able to find the correct information from the first try. It went like Context7 -> Exa -> Supabase -> Cloudflare.
Yeah, but sometimes AI just skips instructions. Since then I re-enabled manual approval to make absolutely sure that tool usage happens only when necessary.
This is a game-changer for AI-driven data workflows. MCP’s standardization removes the friction of connecting tools to live databases, letting data teams query, analyze, and act directly in their AI workspace.
I talk about the latest AI trends and insights. Check out my Substack; you’ll find this deep dive into AI-powered data workflows especially relevant and useful.
That’s so cool Andres! I’ve been building custom ones but this seems to be very easily applicable for any Postgres db.
Sounds like they might also have a version for Claude Code.
Honestly, building them yourself is probably the best and most secure option most of the time, but for databases, this seems like a great alternative.
I personally don’t trust public MCP servers unless they are official or have been vetted by the community.
Which servers are you mostly using, I’m curious?
MCP servers?
I’m most using custom built ones actually, they are for a few different remote db I’m maintaining.
For pre-existing ones, I like file system mcp for Claude; and n8n and playwright for Cursor.
What do you use?
Great article! I love MCPs and use them constantly, but you really have to be careful of how many you use at a time.
It happened to me more than once to have Sonnet call 4 servers one ofter the other as it was not able to find the correct information from the first try. It went like Context7 -> Exa -> Supabase -> Cloudflare.
Other than that they are EXTREMELY useful!
That sounds costly! Have you tried setting guidelines in your AGENTS.md file?
Glad you enjoyed my article by the way
Yeah, but sometimes AI just skips instructions. Since then I re-enabled manual approval to make absolutely sure that tool usage happens only when necessary.
True, it doesn't always follow the instructions correctly
This is a game-changer for AI-driven data workflows. MCP’s standardization removes the friction of connecting tools to live databases, letting data teams query, analyze, and act directly in their AI workspace.
I talk about the latest AI trends and insights. Check out my Substack; you’ll find this deep dive into AI-powered data workflows especially relevant and useful.