Quick Start#

MS-Agent is the official Agent framework launched by the ModelScope community. This framework is committed to using a clear and simple universal capability framework to solve proprietary problems in several domains.

Currently, the domains we are exploring include:

  • DeepResearch: Generate in-depth research reports in the scientific research field

  • CodeGenesis: Generate runnable software project code from requirements

  • General domain: MS-Agent adapts to general LLM conversation scenarios and is compatible with MCP tool calling

MS-Agent is also the backend agent framework for mcp-playground on the ModelScope official website. If developers are interested in the above domains, or hope to learn Agent technology principles and conduct secondary development, welcome to use MS-Agent.

Installation#

For MS-Agent installation, please refer to the installation documentation.

Usage Examples#

The following example can start a general agent conversation:

import asyncio
import sys

from ms_agent import LLMAgent
from ms_agent.config import Config

async def run_query(query: str):
    config = Config.from_task('ms-agent/simple_agent')
    # TODO change to your real api key https://modelscope.cn/my/myaccesstoken
    config.llm.modelscope_api_key = 'xxx'
    engine = LLMAgent(config=config)

    _content = ''
    generator = await engine.run(query, stream=True)
    async for _response_message in generator:
        new_content = _response_message[-1].content[len(_content):]
        sys.stdout.write(new_content)
        sys.stdout.flush()
        _content = _response_message[-1].content
    sys.stdout.write('\n')
    return _content


if __name__ == '__main__':
    query = 'Introduce yourself'
    asyncio.run(run_query(query))

Using Command Line#

ms-agent run --config ms-agent/simple_agent --modelscope_api_key xxx

The above two examples have the same effect and both can conduct multi-turn conversations with the model. Developers can also refer to the following usage methods: