I shipped mcp-probe — a CLI that points at any MCP server, enumerates every tool, resource, and prompt, calls each with auto-generated arguments, validates against declared schemas, prints a pass/fail scorecard, and exits 0/1 for CI.

The plan for launch week: run it against the official Node MCP servers and post results. The first run made me look like I'd broken half the ecosystem. The second, after I read my own output, told a different story — most failures were bugs in my client, not the servers. The rest collapsed into one finding about schema design.

This post is the corrected version. Three sections: what mcp-probe does, what the scorecards say, and the three bugs I fixed in my own client first.

1. What mcp-probe does

One command. stdio, SSE, or Streamable HTTP transport.…