General feedback after a bit of testing

[matthiasgeihs]

I like that this extension exists.
Tab completion worked fine for me in my limited testing.
What didn't work well at all was the chat integration.

• "Llama Agent" doesn't work at all. Just gives "Error"
• "Chat with AI" (without project context) seems utterly useless here. Why would I want a simple chatbot integration without code context in my IDE? There is already so many chatbots UIs out there. This option just clutters the interface.
• "Chat with AI with project context" should be the default chat interface. However, the UI looks weird when I first open it. The conversation history overlays the actual chat window to a point where I thought the chat wasn't even there and something went wrong. Also, a big missing feature is a better editor integration with the chat interface. When I asked it to refactor some code, it simply output the new code to the chat window. There was no easy way to see which changes where made or apply them to the open code window.

I think on a technical level this plugin already works quite well. The responses I get from the AI (tab completions or chat) are already quite useful. However, the UI integration is lacking severely.

[igardev]

@matthiasgeihs , thank you for the honest feedback. It is useful and you have right, the UI needs improvement (and many other things as well). Below are my comments.

• "Llama Agent" doesn't work at all. Just gives "Error": Llama Agent is far from the quality of the agents of some other products and extensions. It is in development, but I prefer to release it and get feedback (like yours). Nevertheless, if it has propper configuration, it produces some results (with intelligent model with tools). You cold use it with local server or with external server. For local server - from llama-vscode menu -> "Select/Start Env..." and select "Local, full package ......" (the apprpriate one for your machine) . This will download the models and start the servers. (Actually llama.cpp will need some more days to fix some issues of gpt-oss with tools, expected in week 18-14, August 2025); For external server - if you have OpenRouter api key, select "Full package - ...." (the apprpriate one for your machine). It will ask you for your OpenRouter key. After this preparation you could write your request and send it to the agent.

• "Chat with AI" (without project context) seems utterly useless here: It is not obligatory to show it if you don't need it. For me it is useful to ask quick simple questions to a local model, without leaving the IDE, instead of searching with google (or use other chatbot UI). Ctrl+; ask question and get an answer, Esc and go back exactly where you left the editor. Another reason is a new feature, which will be released with version 0.0.16 . In this version, there will be features Chat with AI about llama vscode and Local ai runner, which use "Chat with UI" view.

• "Chat with AI with project context" - the UI looks weird: Yes, the conversation history overlays should be fixed. As for the integration with the editor - only the agent can change the source code in the project. "Chat with AI" and "Chat with AI with project context" don't use tools and can't change the source code. On the other hand they could work with smaller and less intelligent local models. Some day probably only the agent will remain, but now it is far from ready and the local models are not yet very intelligent to support it well.

Thanks for spending time to write this feedback. I will have it in mind in the future development.

[matthiasgeihs]

@igardev, thanks for the reply.

1. I selected "Select/start orchestra" -> "12. Agent, completions & chat medium ...". It doesn't work. When I type something into the agent text box and press "Enter", the text disappears and below it shows a green dot and "Error".

2. I definitely see that a simple AI chat is useful. But couldn't you have the same experience integrated with "Chat with AI with project context"? FYI: Personally I wrote myself a tool to have a local LLM always at my fingertips: https://github.com/matthiasgeihs/smart-lookup

Thank you for working on this. I know UI is not easy, but I think AI is actually decent a writing it ;) Overall llama.vscode currently seems to give too many options. I would try to trim it down a bunch. I'd be super keen to have the option to have a completely local and open source alternative to the likes of Windsurf, Cursor, and VSCode GitHub Copilot.