mcp-server-testing

An MCP server's quality metric is agent task success, not endpoint coverage. Unit tests catch code bugs; they say nothing about whether an agent with zero context can pick the right tool and finish a real task. This skill encodes the missing layers: protocol verification (Inspector + conformance) and agent evaluation (verifiable tasks + transcript-driven iteration) — the methodology agents skip, producing servers that "work" and fail in use.

When NOT to use

• Fixing the bugs testing finds (stdout pollution, lifecycle, schema code) → mcp-server-implementation. Rewriting descriptions/surface → mcp-server-design.
• Security checklists/scans → mcp-server-security. Auth flow issues → mcp-server-auth.
• Generic unit testing of non-MCP code, or choosing LLMs.

Workflow