Type: url
Original: https://github.com/Accenture/mcp-bench
Status: generated
Language: en
Summary
MCP-Bench is an open-source benchmark, published by Accenture on GitHub, that evaluates how well large language models discover, select and use tools exposed through the Model Context Protocol to complete realistic tasks. It connects agents to 28 real MCP servers spanning domains such as biomedical research, finance, maps and academic search, runs single-server and multi-server (two- and three-server) task sets, and scores models with a mix of rule-based schema checks and an LLM judge (o4-mini) on task completion, tool usage and planning. The repository includes the task-synthesis pipeline, the multi-round task executor, and a leaderboard, and the work was accepted to a NeurIPS 2025 workshop.