Inquiry: Roadmap for Autonomous Agent Support & Offer to Contribute

[WuJunkai2004]

Hi madox2 and maintainers,

First of all, thank you for building this fantastic plugin! I've been really enjoying using it.

I am very interested in autonomous agent workflows within Vim/Neovim. Looking through the codebase, I noticed that there are already some great foundational pieces in place. For example, the parsing of <<< tool_call and <<< tool_response in utils.py, as well as the support for MCP via third-party providers.

I wanted to ask about the current progress or future roadmap for first-class agent support (e.g., allowing the plugin to autonomously execute tool calls and feed the responses back into the context loop without manual intervention).

If this is a feature you are planning or open to adding, I would love to contribute to the effort! Please let me know if there are any ongoing design discussions, or if there are specific areas/PRs where I could start helping out.

Thanks again for the great work!

[madox2]

Hi, thanks for opening the discussion!

MCP support has been already discussed in #159, it ended up creating 3rd party mcp provider. That's the preferred way for now.

Adding MCP to core vim-ai plugin would mean to implement MCP client from scratch (I want this plugin to stay zero dependencies and platform independent) and provide solid permission system to control running tools. If you are interested in implementing it, I would suggest to start with a custom provider and will see where it goes. If there are any missing pieces in core plugin, I will be glad to add it.

When it comes to full agentic support, there are better (and more complex) tools out there, like claude code, that will vim-ai never be able to compete with. I see vim-ai as a quick tool that interacts with llm and therefore I would like to keep things rather simple.

I was recently thinking of adding a mechanism that allows user to define custom functions/tools that delegates calls to vim functions. That could be used to define custom tools in a simple way (e.g. a tool that retrieves filetype or the content of the current file), but maybe it would be also possible somehow hook it into MCP. If you like the idea I can expand on that.

[WuJunkai2004]

I thought that a simple MCP system which AI can call tools both builtin or MCP extensions is not complex. I can write it in no more than 200 lines python codes without 3rd parth modules.
And yes, I saw the #159. It make a foundation for agent, and we can devlop it better.

If you have intersting, I can submit a PR just use subprocess and threading.
I am writing this issue because I regard vim-ai as an equivalent to the Copilot plugin on VSCode, so I hope they can be equally powerful.

[madox2]

Sounds great, feel free to open PR, we can discuss details there

[WuJunkai2004]

#206 the PR is submitted

[madox2]

Hi, I had a long thought about adding agentic capabilities to the vim-ai and decided to step back on it.
One reason is that even though it would be nice to have MCP, skills, etc. support, but this would never become a SOTA tool for the agentic workfows.
And on the other hand, it would require a significant amount of time to implement and maintain these functionalities (even more without using any 3rd party packages).
Therefore, I would rather focus on simple and seamless LLM integration in vim and do it right.

I believe that if anybody is interested in implementing agentic features into vim-ai, it should be possible through a custom provider.
If vim-ai plugin lacks anything to accomplish that, I am open to add it.

@WuJunkai2004 and others, I truly appreciate your work and interest in this, but I need to close this issue.