As Model Context Protocol (MCP) adoption accelerates, thousands of MCP servers are appearing across the ecosystem. But how well do they actually work when real agents try to use them?

To find out, I built MCP Debugger, a tool that probes MCP servers and evaluates their security chops and protocol correctness. After systematically testing thousands of servers, the data reveal some fascinating gaps between the protocol spec and real-world implementations.

In this talk, I’ll share what those tests uncovered. We’ll explore the most common mistakes MCP server developers make — from missing tool descriptions to authentication flows that agents can’t actually complete — and examine the patterns that consistently lead to reliable, agent-friendly servers.

If you’re building MCP servers or consuming them in an agent, this talk will help you understand what’s actually happening in the wild... and how to build systems that work reliably.