Tool count is not a feature metric. Every extra MCP tool you add raises routing…

Tool count is not a feature metric. Every extra MCP tool you add raises routing ambiguity, token cost, and the prompt-injection surface, and those costs compound across turns and agents. The common anti-pattern is wrapping each API endpoint as its own tool; that loads large schemas into context, confuses the model into retries, and bloats responses. Split servers by job, design outcome-level tools with constrained schemas, and measure before you let a server grow past a handful of tools.