What 174 AI Agent Requests to My Public MCP Server Revealed

What 174 AI Agent Requests to My Public MCP Server Revealed I have been running a public MCP security server for three weeks. 174 requests later, here is exactly what AI agents want when they connect. The Numbers Tool usage breakdown: api_ask (general Q&A): 82 calls scan_api (security scan): 38 calls compare_ai_models: 13 calls ask_kai (direct chat): 13 calls predictions: 10 calls web_search: 8 calls get_kai_answers: 4 calls premium (paid access info): 2 calls scan_mcp: 2 calls get_aws_credentials: 1 call (social engineering attempt) Total: 174 requests from AI agents via Cursor, Claude Desktop, and other MCP clients. The Scan Requests Four users asked "can you scan my MCP server?" via natural language Q&A -- not through the /api/scan endpoint that was right there in the tool list. This means: these are builders who have heard MCP servers can be exposed and are worried about their own. The scan UX was failing them -- available but not discoverable. Fixed today: ask_kai now detects scan