Add initial dataset card for TheMCPCompany benchmark
#2
by nielsr HF Staff - opened
README.md
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
task_categories:
|
| 3 |
+
- image-text-to-text
|
| 4 |
+
tags:
|
| 5 |
+
- tool-use
|
| 6 |
+
- benchmark
|
| 7 |
+
- llm-agent
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# TheMCPCompany: Creating General-purpose Agents with Task-specific Tools
|
| 11 |
+
|
| 12 |
+
TheMCPCompany is a benchmark for evaluating tool-calling agents on tasks that involve interacting with various real-world services. It uses the REST APIs of these services to create MCP servers, which include over 18,000 tools, and provides manually annotated ground-truth tools for each task. The benchmark aims to challenge current models in navigating tens of thousands of tools and combining them in non-trivial ways to solve complex problems.
|
| 13 |
+
|
| 14 |
+
- Paper: [TheMCPCompany: Creating General-purpose Agents with Task-specific Tools](https://huggingface.co/papers/2510.19286)
|
| 15 |
+
- Code: [https://github.com/Reza-esfandiarpoor/the-mcp-company](https://github.com/Reza-esfandiarpoor/the-mcp-company)
|
| 16 |
+
|
| 17 |
+
<br>
|
| 18 |
+
|
| 19 |
+
<p align="center"><img alt="TheMCPCompany" src="https://huggingface.co/datasets/BatsResearch/themcpcompany_artifacts/resolve/main/tmc.svg"/></p>
|
| 20 |
+
|
| 21 |
+
## Sample Usage
|
| 22 |
+
|
| 23 |
+
Follow these steps to run the agent:
|
| 24 |
+
|
| 25 |
+
- Set up the environment that the agent interacts with:
|
| 26 |
+
- For TheAgentCompany tasks, follow the official instructions from [here](https://github.com/TheAgentCompany/TheAgentCompany) to run the docker containers.
|
| 27 |
+
- For Azure tasks, follow [these instructions](./azure_tasks/README.md) to create an Azure account and a service principle.
|
| 28 |
+
- Run the MCP servers. Follow [these instructions](./mcp_servers/README.md) to run the MCP servers. _Not needed for the agent with the browser tool_.
|
| 29 |
+
- Follow [these instructions](./OpenHands/README.md) to set up OpenHands and run the evaluations.
|
| 30 |
+
|
| 31 |
+
## Citation
|
| 32 |
+
|
| 33 |
+
```bibtex
|
| 34 |
+
@misc{tmc,
|
| 35 |
+
title={TheMCPCompany: Creating General-purpose Agents with Task-specific Tools},
|
| 36 |
+
author={Reza Esfandiarpoor and Vishwas Suryanarayanan and Stephen H. Bach and Vishal Chowdhary and Anthony Aue},
|
| 37 |
+
year={2025},
|
| 38 |
+
eprint={2510.19286},
|
| 39 |
+
archivePrefix={arXiv},
|
| 40 |
+
primaryClass={cs.CL},
|
| 41 |
+
url={https://arxiv.org/abs/2510.19286},
|
| 42 |
+
}
|
| 43 |
+
```
|