mcp-tool-bench/MCPToolBenchPP

Comprehensive benchmark for evaluating LLM tool-use capabilities across 45+ MCP server categories (browser automation, file systems, search, maps, payments, finance) with 4k+ instances covering single and multi-step tool calls. Evaluation uses standardized metrics (AST and Pass@K) with an LLM-as-judge approach, supporting major models like GPT-4o, Qwen, and Claude across multilingual scenarios. Integrates with MCP ecosystem servers and the OneKey MCP Router for simplified API access to commercial services like Google Maps and Perplexity.