<h1 align="center">
🦊 MCPBench: A Benchmark for Evaluating MCP Servers
</h1>

<div align="center">

[![Documentation][docs-image]][docs-url]
[![Package License][package-license-image]][package-license-url]

</div>

<div align="center">
<h4 align="center">

中文 |
English

</h4>
</div>

MCPBench is an evaluation framework for MCP Servers. It supports the evaluation of three types of servers: Web Search, Database Query and GAIA, and is compatible with both local and remote MCP Servers. The framework primarily evaluates different MCP Servers (such as Brave Search, DuckDuckGo, etc.) in terms of task completion accuracy, latency, and token consumption under the same LLM and Agent configurations. Here is the evaluation report.

> The implementation refers to LangProBe: a Language Programs Benchmark.\
> Big thanks to Qingxu Fu for the initial implementation!

<hr>

📋 Table of Contents

- 🔥 News
- 🛠️ Installation
- 🚀 Quick Start
- Launch MCP Server
- Launch Evaluation
- 🧂 Datasets and Experiments
- 🚰 Cite

🔥 News

+ Sep. 1, 2025 🌟 Modelscope AI hackathon will be hold on Sep. 23rd, ref: https://modelscope.cn/active/aihackathon-mcp-agent + Apr. 29, 2025 🌟 Update the code for evaluating the MCP Server Package within GAIA. + Apr. 14, 2025 🌟 We are proud to announce that MCPBench is now open-sourced.

🛠️ Installation

The framework requires Python version >= 3.11, nodejs and jq.

conda create -n mcpbench python=3.11 -y
conda activate mcpbench
pip install -r requirements.txt

🚀 Quick Start

Please first determine the type of MCP server you want to use: - If it is a remote host (accessed via SSE, such as ModelScope, Smithery, or localhost), you can directly conduct the evaluation. - If it is started locally (accessed via npx using STDIO), you need to launch it.

Launch MCP Server (optional for stdio)

First, you need to write the following configuration:

{
    "mcp_pool": [
        {
            "name": "firecrawl",
            "run_config": [
                {
                    "command": "npx -y firecrawl-mcp",
                    "args": "FIRECRAWL_API_KEY=xxx",
                    "port": 8005
                }
            ]
        }  
    ]
}

Save this config file in the configs folder and launch it using:

sh launch_mcps_as_sse.sh YOUR_CONFIG_FILE

For example, save the above configuration in the configs/firecrawl.json file and launch it using:

sh launch_mcps_as_sse.sh firecrawl.json

Launch Evaluation

To evaluate the MCP Server's performance, you need to set up the necessary MCP Server information. the code will automatically detect the tools and parameters in the Server, so you don't need to configure them manually, like:

{
    "mcp_pool": [
        {
            "name": "Remote MCP example",
            "url": "url from https://modelscope.cn/mcp or https://smithery.ai"
        },
        {
            "name": "firecrawl (Local run example)",
            "run_config": [
                {
                    "command": "npx -y firecrawl-mcp",
                    "args": "FIRECRAWL_API_KEY=xxx",
                    "port": 8005
                }
            ]
        }  
    ]
}

To evaluate the MCP Server's performance on WebSearch tasks:

sh evaluation_websearch.sh YOUR_CONFIG_FILE

To evaluate the MCP Server's performance on Database Query tasks:

sh evaluation_db.sh YOUR_CONFIG_FILE

To evaluate the MCP Server's performance on GAIA tasks:

sh evaluation_gaia.sh YOUR_CONFIG_FILE

For example, save the above configuration in the configs/firecrawl.json file and launch it using:

sh evaluation_websearch.sh firecrawl.json

Datasets and Experimental Results

Our framework provides two datasets for evaluation. For the WebSearch task, the dataset is located at MCPBench/langProBe/WebSearch/data/websearch_600.jsonl, containing 200 QA pairs each from Frames, news, and technology domains. Our framework for automatically constructing evaluation datasets will be open-sourced later.

For the Database Query task, the dataset is located at MCPBench/langProBe/DB/data/car_bi.jsonl. You can add your own dataset in the following format:

{
  "unique_id": "",
  "Prompt": "",
  "Answer": ""
}

We have evaluated mainstream MCP Servers on both tasks. For detailed experimental results, please refer to Documentation

🚰 Cite

If you find this work useful, please consider citing our project or giving us a 🌟:

@misc{mcpbench,
  title={MCPBench: A Benchmark for Evaluating MCP Servers},
  author={Zhiling Luo, Xiaorong Shi, Xuanrui Lin, Jinyang Gao},
  howpublished = {\url{https://github.com/modelscope/MCPBench}},
  year={2025}
}

Alternatively, you may reference our report.

@article{mcpbench_report,

      title={Evaluation Report on MCP Servers}, 

      author={Zhiling Luo, Xiaorong Shi, Xuanrui Lin, Jinyang Gao},

      year={2025},

      journal={arXiv preprint arXiv:2504.11094},

      url={https://arxiv.org/abs/2504.11094},

      primaryClass={cs.AI}

}

[docs-image]: https://img.shields.io/badge/Documentation-EB3ECC
[docs-url]: https://arxiv.org/abs/2504.11094
[package-license-image]: https://img.shields.io/badge/License-Apache_2.0-blue.svg
[package-license-url]: https://github.com/modelscope/MCPBench/blob/main/LICENSE