LiveMCPBench: Can Agents Navigate an Ocean of MCP Tools?

Benchmarking the agent in real-world tasks within a large-scale MCP toolset.

🌐 Website   |   📄 Paper   |   🤗 Dataset   |   🐳 Docker   |   🏆 Leaderboard   |   🙏 Citation

News

• [8/18/2025] We releas Docker images and add evaluation results in leaderboard for three new models: GLM 4.5, GPT-5-Mini, and Kimi-K2.
• [8/3/2025] We release the LiveMCPBench.

Getting Started

Prerequisites

We recommend using our docker image, but if you want to run the code locally, you will need to install the following tools:

• npm
• uv

Installation

1. Pull the docker image

   docker pull hysdhlx/livemcpbench:latest

2. Git the repo and run the docker image

   git clone https://github.com/icip-cas/LiveMCPBench.git
   cd LiveMCPBench
   
   docker run -itd \
   -v "$(pwd):/outside" \
   --gpus all \
   --ipc=host \
   --net=host \
   --name LiveMCPBench_container \
   hysdhlx/livemcpbench:latest \
   bash

3. Prepare the .env file

   cp .env_template .env

   You can modify the .env file to set your own environment variables.

   # MCP Copilot Agent Configuration
    BASE_URL=
    OPENAI_API_KEY=
    MODEL=
   
    # Tool Retrieval Configuration
    EMBEDDING_MODEL=
    EMBEDDING_BASE_URL=
    EMBEDDING_API_KEY=
    EMBEDDING_DIMENSIONS=1024
    TOP_SERVERS=5
    TOP_TOOLS=3
    # Abstract API Configuration (optional)
    ABSTRACT_MODEL=
    ABSTRACT_API_KEY=
    ABSTRACT_BASE_URL=
   
    # Proxy Configuration (optional)
    http_proxy=
    https_proxy=
    no_proxy=127.0.0.1,localhost
    HTTP_PROXY=
    HTTPS_PROXY=
    NO_PROXY=127.0.0.1,localhost
   
    # lark report (optional)
    LARK_WEBHOOK_URL=

4. Enter the container & Reset the environment

   As we have mounted the code repo to /outside, you can access the code repo in the container at /outside/.

   docker exec -it LiveMCPBench_container bash

   Because the agent may change the environment, we recommend resetting the environment before running the agent. To reset the environment, you can run the following command:

   cd /LiveMCPBench/
   bash scripts/env_reset.sh 

   This will copy the repo code in /outside to /LiveMCPBench and link the annotated_data to /root/.

5. Check the MCP tools

   bash ./tools/scripts/tool_check.sh

   After running this command, you can check ./tools/test/tools.json to see the tools.

   You could run this script multiple times if you find some tools are not working.

6. Index the servers

   The MCP Copilot Agent requires you have indexed the servers before running. You can run the following command to warm up the agent:

   uv run -m baseline.mcp_copilot.arg_generation

Quick Start

MCP Copilot Agent

Example Run

bash ./baseline/scripts/run_example.sh

This will run the agent with a simple example and save the results in ./baseline/output/.

Full Run

We default use /root dir to store our data that the agent will access. If you want to run locally, you need to ensure the file in the right path.

1. Run the MCP Copilot Agent

   Be sure you have set the environment variables in the .env file.

   bash ./baseline/scripts/run_baselines.sh

2. Check the results

   After running the agent, you can check the trajectories in ./baseline/output.

Evaluation using the LiveMCPEval

1. Modify the MODEL in .env to change evluation models

2. Run the evaluation script

   bash ./evaluator/scripts/run_baseline.sh

3. Check the results

   After running the evaluation, you can check the results in ./evaluator/output.

4. Calculate the success rate

   uv run ./evaluator/stat_success_rate.py --result_path /path/to/evaluation/

Project Structure

LiveMCPBench/
├── annotated_data/      # Tasks and task files
├── baseline/            # MCP Copilot Agent
│   ├── scripts/         # Scripts for running the agent
│   ├── output/          # Output for the agent
│   └── mcp_copilot/     # Source code for the agent
├── evaluator/           # LiveMCPEval
│   ├── scripts/         # Scripts for evaluation
│   └── output/          # Output for evaluation
├── tools/               # LiveMCPTool
│   ├── LiveMCPTool/     # Tool data
│   └── scripts/         # Scripts for the tools
├── scripts/             # Path prepare scripts
├── utils/               # Utility functions
└── .env_template        # Template for environment

Citation

If you find this project helpful, please use the following to cite it:

@misc{mo2025livemcpbenchagentsnavigateocean,
      title={LiveMCPBench: Can Agents Navigate an Ocean of MCP Tools?}, 
      author={Guozhao Mo and Wenliang Zhong and Jiawei Chen and Xuanang Chen and Yaojie Lu and Hongyu Lin and Ben He and Xianpei Han and Le Sun},
      year={2025},
      eprint={2508.01780},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2508.01780}, 
}