2025-10-28: We released our paper and project page! 🎉

📄 Read the Paper  |  🌐 Visit the Project Page

OSWorld-MCP is a comprehensive and fair benchmark for evaluating computer-use agents in real-world scenarios.
It jointly measures Model Context Protocol (MCP) tool invocation capabilities, graphical user interface (GUI) operation skills, and decision-making performance.
Designed as an extension of OSWorld, it significantly improves realism, balance, and comparability in evaluation.

Key Features & Findings

• 158 validated MCP tools, spanning 7 common applications (LibreOffice Writer, Calc, Impress, VS Code, Google Chrome, VLC, OS utilities). Among them, 25 distractor tools for robustness testing
• 250 tool-beneficial tasks → 69% of benchmark tasks benefit from MCP tools
• Multi-round tool invocation possible, posing real decision-making challenges
• MCP tools boost model accuracy & efficiency — e.g., OpenAI o3: 8.3% → 20.4% (15 steps)
• Highest observed Tool Invocation Rate (TIR) = 36.3% (Claude-4-Sonnet, 50 steps) → indicating ample room for improvement
• MCP tools improve agent metrics
• Higher tool invocation correlates with higher accuracy
• Combining tools introduces significant challenges

Architecture Overview

Figure: OSWorld-MCP evaluation framework integrating GUI actions and MCP tool invocations.

# Clone OSWorld base repo
git clone https://github.com/xlang-ai/OSWorld.git

# Clone OSWorld-MCP
git clone https://github.com/X-PLUG/OSWorld-MCP.git

Integrate OSWorld-MCP files into OSWorld to enable MCP support.

1. Copy MCP files into /home inside Docker:

/home/
└── mcp_server/
└── osworld_mcp_client.py

1. Install dependencies:

pip install -r requirements.txt

1. Install Node.js
2. Launch MCP server:

cd mcp_server
bash debug_server.sh

A successful launch opens the local MCP debug UI in your browser.

Example: Evaluate Claude 4 Sonnet (15 steps):

python run_multienv_e2e.py \
    --api_url <your_api_url> \
    --api_key <your_api_key> \
    --model 'claude-sonnet-4-20250514-thinking' \
    --test_all_meta_path 'evaluation_examples/test_all.json' \
    --num_envs 1 \
    --action_space mcp \
    --max_steps 15 \
    --max_trajectory_length 15

1. Task Accuracy (Acc) — % of tasks successfully completed.
2. Tool Invocation Rate (TIR) — correct decisions to use a tool or not.
3. Average Completion Steps (ACS) — average number of actions per completed task.

🔗 Live Leaderboard: osworld-mcp.github.io

Max Steps: 15

Max Steps: 50

@article{jia2025osworldmcp,
  title={OSWorld-MCP: Benchmarking MCP Tool Invocation in Computer-Use Agents},
  author={Jia, Hongrui and Liao, Jitong and Zhang, Xi and Xu, Haiyang and Xie, Tianbao and Jiang, Chaoya and Yan, Ming and Liu, Si and Ye, Wei and Huang, Fei},
  year={2025},
  journal={arXiv preprint arXiv:2510.24563}
}