MCP Server Testing Suite

This skill provides two testing modes for MCP servers, both powered by @mcp-use/inspector.

How it works

The inspector runs a local web server that proxies JSON-RPC requests to your MCP server. Every MCP operation (initialize, tools/list, tools/call, resources/list, prompts/get, etc.) goes through curl calls to the inspector's API. This means you can test any MCP server reachable over HTTP/SSE or WebSocket without needing a config file or daemon.

Command: `/mcp-test`

Basic protocol-level testing. Verifies every MCP primitive works correctly.

When the user invokes /mcp-test, follow the guide in references/basic-test-guide.md. This command:

Asks the user for their MCP server URL (or detects it from context)
Starts the inspector in the background
Runs through every MCP primitive systematically
Reports pass/fail for each check
Cleans up the inspector process

Read references/basic-test-guide.md for the complete step-by-step procedure.

Command: `/mcp-test-llm`

End-to-end LLM-powered testing. Uses a real LLM to generate business-relevant test scenarios and execute them against the MCP server with tool-use.

When the user invokes /mcp-test-llm, follow the guide in references/llm-test-guide.md. This command:

Asks the user for their MCP server URL
Asks for an LLM API key (OpenAI, Anthropic, Google, or OpenRouter)
Optionally saves credentials to .env
Discovers all tools and their schemas
Generates business-relevant test cases based on what the tools actually do
Executes each test case through the LLM chat endpoint
Validates tool calls were made and results are sensible
Reports comprehensive results

Read references/llm-test-guide.md for the complete step-by-step procedure. Read references/providers.md for LLM provider configuration details (including OpenRouter).

Reference files

File	When to read
`references/basic-test-guide.md`	When running `/mcp-test`
`references/llm-test-guide.md`	When running `/mcp-test-llm`
`references/inspector-api.md`	Full API reference for the inspector (both commands use this)
`references/providers.md`	LLM provider setup including OpenRouter, base URL overrides
`references/troubleshooting.md`	When something fails and you need to debug
`references/business-cases.md`	How to generate realistic business test cases from tool schemas

Important conventions

Always use --no-open when starting the inspector (agents don't need browsers)
Always disable telemetry: MCP_USE_ANONYMIZED_TELEMETRY=false
Always clean up the inspector process when done (kill $INSPECTOR_PID)
Use the crash thinking tool to structure your reasoning at each testing phase
Report results as a clear pass/fail table, not walls of JSON
If a test fails, dig into why before moving on — read error messages, check schemas, retry with different inputs

mcp-server-tester