Framework or libraries can be used in this project

[MuggleJinx]

For this project, we need the agent level's ability to analyze code repository and documents (readme, cookbooks, etc) to detect useful APIs.

We also need to manage the running environment of different projects. Welcome to share your ideas!

[MuggleJinx]

Some examples might be useful are:

• Composio's SWE Kit: https://composio.dev/swe-kit/#usecases
• Describer : https://github.com/apideck-libraries/describer
• Codebase Digest: https://github.com/kamilstanuch/codebase-digest
• Project Guide: https://github.com/sojohnnysaid/project-guide

[qingshungLI]

Hello, Senior Xiaotian. I am Espier from the WeChat group. I watched the video of the Tencent meeting that day again, and tried it myself for a while. I also explored several links in your issue. Let me share my superficial thoughts:
1 Positioning: There are many similar agents on the market, whether open source or closed source, basically need to manually set the MCP, and there are obvious problems such as insufficient MCP and imperfect retrieval. I think this is where we should improve. Let the multi-agent framework help users set the MCP, and mcpify the open source project on GitHub, which can basically solve the above two problems.
2. Task or repo driven: I think it should be repo driven (find valuable projects on GitHub and mcpify them). Of course, when the repo's mcp cannot meet the user's needs, you can do something similar to deepwiki, that is, the user asks the agent to find the project immediately and mcpify it.
3. How to detect projects: From my understanding, this is not a difficult point under the repo drive. The only difficult thing may be that the token of LLM is not long enough, which can be solved by the agent framework. From the three github open source projects you posted, it seems that AsyncFuncAI/deepwiki-open: Open Source DeepWiki: AI-Powered Wiki Generator for GitHub/Gitlab/Bitbucket Repositories. Join the discord: https://discord.gg/gMwThUMeme has similar effects, but the generated content is a bit long. I tried it and wanted to use mcpify to let AI understand and save it, but it was not suitable. It is better to write a package that generates concise and formatted content that can be understood by LLM (I think the understanding content generated by deepwiki is very good, and the cli.py in the current code base is far from enough)
4. Manage different environments: Open source projects can be solved with Docker. Running open source projects on Docker only requires a few lines of command lines, and can also be made into a web page and deployed on the server.
5. Difficulties in my opinion: Although many agents are now using the MCP brand, there is still a long way to go in terms of automatic configuration, server richness, and calling flexibility. I think the key to the project is how to find good projects on GitHub. In the initial stage, you can manually search and mcpify according to the number of stars. As the tasks gradually become richer, how to find the most suitable and effective projects on GitHub is very important. This is also the problem that AI search needs to solve. I feel that Perplexity's Pro Search is more useful and can be used to search for suitable projects.

The above is my recent idea. It is very easy to make an agent that includes MCP, and there are many on the market. However, there are very few agents that can surpass the commonly used large models and truly enable MCP to benefit daily work. I am full of interest and passion for this. Welcome to communicate, and I also hope to be able to participate in the advancement of the project!

[MuggleJinx]

@qingshungLI hi Li! Thanks for the detailed discussion! Your ideas are very insightful! I agreed with u that we can start with projects with many stars.
Sure, feel free to contribute and implement your ideas! This project just starts and we can help define the roadmaps for the future developments (I am currently on vacation leave, but I will find time and consult you). Thank you again!