Preprint · 2025

MCPMark: A Benchmark for Stress-Testing Realistic and Comprehensive MCP Use

Zijian Wu, Xiangyan Liu, Xinyuan Zhang, Lingjun Chen, Fanqing Meng, Lingxiao Du, Yiran Zhao, Fanshi Zhang, Yaoqi Ye, Jiawei Wang, Zirui Wang, Jinjie Ni, Yufan Yang, Arvin Xu, Michael Qizhe Shieh

Abstract

MCP standardizes how LLMs interact with external systems, forming the foundation for general agents. However, existing MCP benchmarks remain narrow in scope: they focus on read-heavy tasks or tasks with limited interaction depth, and fail to capture the complexity and realism of real-world workflows. To address this gap, we propose MCPMark, a benchmark designed to evaluate MCP use in a more realistic and comprehensive manner. It consists of $127$ high-quality tasks collaboratively created by domain experts and AI agents. Each task begins with a curated initial state and includes a programmatic script for automatic verification. These tasks demand richer and more diverse interactions with the environment, involving a broad range of create, read, update, and delete (CRUD) operations. We conduct a comprehensive evaluation of cutting-edge LLMs using a minimal agent framework that operates in a tool-calling loop. Empirical results show that the best-performing model, gpt-5-medium, reaches only $52.56$\% pass@1 and $33.86$\% pass^4, while other widely regarded strong models, including claude-sonnet-4 and o3, fall below $30$\% pass@1 and $15$\% pass^4. On average, LLMs require $16.2$ execution turns and $17.4$ tool calls per task, significantly surpassing those in previous MCP benchmarks and highlighting the stress-testing nature of MCPMark.

BibTeX Copy

@misc{wu_mcpmark_2025,
	title = {{MCPMark}: {A} {Benchmark} for {Stress}-{Testing} {Realistic} and {Comprehensive} {MCP} {Use}},
	shorttitle = {{MCPMark}},
	url = {http://arxiv.org/abs/2509.24002},
	doi = {10.48550/arXiv.2509.24002},
	abstract = {MCP standardizes how LLMs interact with external systems, forming the foundation for general agents. However, existing MCP benchmarks remain narrow in scope: they focus on read-heavy tasks or tasks with limited interaction depth, and fail to capture the complexity and realism of real-world workflows. To address this gap, we propose MCPMark, a benchmark designed to evaluate MCP use in a more realistic and comprehensive manner. It consists of \$127\$ high-quality tasks collaboratively created by domain experts and AI agents. Each task begins with a curated initial state and includes a programmatic script for automatic verification. These tasks demand richer and more diverse interactions with the environment, involving a broad range of create, read, update, and delete (CRUD) operations. We conduct a comprehensive evaluation of cutting-edge LLMs using a minimal agent framework that operates in a tool-calling loop. Empirical results show that the best-performing model, gpt-5-medium, reaches only \$52.56\${\textbackslash}\% pass@1 and \$33.86\${\textbackslash}\% pass{\textasciicircum}4, while other widely regarded strong models, including claude-sonnet-4 and o3, fall below \$30\${\textbackslash}\% pass@1 and \$15\${\textbackslash}\% pass{\textasciicircum}4. On average, LLMs require \$16.2\$ execution turns and \$17.4\$ tool calls per task, significantly surpassing those in previous MCP benchmarks and highlighting the stress-testing nature of MCPMark.},
	urldate = {2026-11-06},
	publisher = {arXiv},
	author = {Wu, Zijian and Liu, Xiangyan and Zhang, Xinyuan and Chen, Lingjun and Meng, Fanqing and Du, Lingxiao and Zhao, Yiran and Zhang, Fanshi and Ye, Yaoqi and Wang, Jiawei and Wang, Zirui and Ni, Jinjie and Yang, Yufan and Xu, Arvin and Shieh, Michael Qizhe},
	month = sep,
	year = {2025},
	note = {arXiv:2509.24002 [cs.CL]},
	keywords = {Computer Science - Artificial Intelligence, Computer Science - Computation and Language},
}