Hello! I'm the maker of mcpspan. Once an MCP server is live, its users are AI agents, and nothing tells you how they actually use it: which tools they pick, where they get stuck, what they ask for that isn't there. mcpspan answers that. One line in your server, and the dashboard shows: - tool calls, resource reads and prompt gets per client (Claude, Claude Code, Cursor, ChatGPT, VS Code and others), with latency per tool - failures split by kind: a handler that threw, a tool that returned an error, arguments refused before the handler ran (and which argument) - tools agents asked for that your server doesn't have, with the closest name it does have - sessions step by step, calls an agent repeated with the same arguments, and what came right before a failure - what your tool list costs agents in tokens, and which tools they never use Privacy: parameter values never leave your server's process, and error messages can be turned off. SDKs for TypeScript, Python, Go, C#, Java, Rust, Ruby and PHP, all checked against one written contract by the same test suite. It's open source (MIT) and free to self-host with every feature: https://github.com/mcpspan/mcpspan If you'd rather not run it, mcpspan Cloud hosts it for you, with a 14-day trial and no card. Most features came from feedback on the first release, so I'd love to hear what you'd want to know about how agents use your server.

mcpspan

Analytics for MCP servers: see how AI agents use your tools

Analytics for MCP servers: see how AI agents use your tools