MCP Discovery: Why Your MCP Server Needs Better Tool Discovery Than You Think (After 85 Production Outages)

Building MCP servers reveals critical stability issues. After 85 production outages, it became evident that poor tool discovery can compromise an MCP server more severely than bad code, specifically regarding how AI clients identify tools. Furthermore, following 86 outages, random connection drops were traced back to SSE streaming issues. Despite implementing logging, health checks, fuzzy matching for tool discovery, authentication, and rate limiting, the SSE streaming mechanism remained the primary cause of these unexpected server disconnections.