Logo of Model Context Protocol: Model Context Protocol, Public domain, via Wikimedia Commons · Background: David Bedard, Public domain, via Wikimedia Commons

We connected AI agents to our internal tools using MCP. They now read tickets, test the app, and check the database on their own.

Powerful. But in utility billing, one wrong write means a wrong bill in thousands of homes.

4 rules I keep in mind as a tech lead:

1. Treat every MCP server like a dependency
It's code running with your credentials. Review it, pin the version, keep it away from production access.

2. Treat tool output as untrusted input
A ticket or web page can carry hidden instructions. Validate what comes back, like user input.

3. Start read-only
Add writes one tool at a time, with low-privilege credentials. Delete, deploy, send → human approval.

4. Fewer tools, sharper descriptions
Too many tools confuse the model, and it picks tools by their description. Write them like API docs.

MCP hasn't removed integration risk.
It's moved it — from "how do I connect this?"
to "what should this AI never touch?"

Next, I'm thinking of writing about tools that scan MCP servers for hidden risks.

Would that be useful to you?

#MCP #AIAgents #TechLeadership #BuildingInPublic

Originally posted on LinkedIn. Join the discussion there ↗