ModelScope Agents: Open‑Source LLM Agent Framework and Practical Guide

OpenAI DEV Day announced GPT Stores, allowing users to customize GPTs with knowledge bases, web‑browsing, text‑to‑image, and code‑interpreter tools, but developers reported issues such as black‑box behavior, limited tool integration, privacy concerns, and high token costs.

To offer an open‑source alternative, ModelScope Agents was created, enabling users to build and customize agents with any LLM, connect to ModelScope’s model ecosystem, and foster application‑level innovation.

Key advantages of ModelScope Agents:

1. Open source: All code is publicly available for customization and further development.

2. Multi‑LLM support: Users can compare and switch between different LLM agents to find the best cost‑performance balance.

3. Fine‑tuning capability: Provides prompt‑based customization as well as professional fine‑tuning solutions and datasets for domain‑specific agents.

One‑page overview of ModelScope Agents

Practical case

Colab notebook: https://colab.research.google.com/github/modelscope/modelscope-agent/blob/master/demo/modelscope_agents.ipynb

Notebook: https://github.com/modelscope/modelscope-agent/blob/master/demo/demo_modelscopegpt_agent.ipynb

Python environment

Python >= 3.10 (recommended 3.10)

Installation and usage

git clone https://github.com/modelscope/modelscope-agent.git
# Install basic dependencies
pip3 install -r requirements.txt
cd modelscope-agent/apps/agents
# Install app-specific dependencies
pip3 install -r requirements.txt

# Recommended LLM: Tongyi Qianwen 2.0 (DashScope)
# Obtain DASHSCOPE_API_KEY from the DashScope console
export DASHSCOPE_API_KEY=YOUR_KEY_HERE
python3 app.py

RPG game example: Jin Yong’s Martial Arts World

The demo builds an interactive online RPG where players assume a key character from Jin Yong’s novels, with scene descriptions, generated images, and three-choice interactions, all presented in Simplified Chinese.

Roadmap

✓ Support manual configuration of agents

✓ Build agents via LLM dialogue

○ Production‑grade features: token usage stats, runtime logs, performance analysis

○ Effective evaluation loops for iterative model improvement

○ Integration with ModelScope workspace

○ Knowledge‑base retrieval optimization

○ Agent publishing and sharing

○ Support for open‑source models such as ChatGLM, Baichuan, and commercial APIs

○ Long‑text handling in memory

○ Production‑grade logging and performance analysis

○ Agent fine‑tuning

○ Multi‑scenario agent effectiveness evaluation

We welcome developers to contribute via the ModelScope GitHub repository or join our WeChat group for collaboration.

GitHub: https://github.com/modelscope/modelscope-agent