Dive AI is great but have some issues or need some improvement.

[onigetoc]

First of all, your Electron application is extraordinary. Being able to add many models is very useful.
Please continue to make it better.

Some bugs and suggestions.

When we open a new conversation, if there is already something written in the message box and I click on tools, This shows the tools and it's correct, but the message is sent and it shouldn't do that.

When we add a new server and we activate or deactivate it, it seems that all the tools, that all the servers reload. And they should only reload the server, choose and not all the servers. They should, like Cline or Roo Code, have a reload button that only reloads one server. Because it's very long to wait.

Allow deleting server without to have to edit the json file.

I can not see the models names, the menu is not width enough.

i do not know witch one i'm using

Question: I'm curious to know how you can display the entire list of all available models. I was also surprised to have the entire list of Groq models because it use the Open AI compatible.

Questions 2: Where are the server installed? I looked for them in my computer, but I can't find them. (Windows 10)

[johnfunmula]

Thank you for your feedback and suggestions. We will fix these issues in the upcoming few versions.

Question: I'm curious to know how you can display the entire list of all available models. I was also surprised to have the entire list of Groq models because it use the Open AI compatible.

Open AI compatible will directly use the official Open AI SDK to retrieve the model list.

Questions 2: Where are the server installed? I looked for them in my computer, but I can't find them. (Windows 10)

The server is currently packaged within the application, but we have plans to separate it out."

[kevinwatt]

When we add a new server and we activate or deactivate it, it seems that all the tools, that all the servers reload. And they should only reload the server, choose and not all the servers. They should, like Cline or Roo Code, have a reload button that only reloads one server. Because it's very long to wait.

We are currently converting the MCP Host component from TypeScript to Python. At the time, the MCP Reload was designed to conveniently update everything at once during activation. After the MCP Host migration is complete, we will modify the MCP Loading mechanism to operate individually

Allow deleting server without to have to edit the json file.

This feature will be developed and updated next week

Question: I'm curious to know how you can display the entire list of all available models. I was also surprised to have the entire list of Groq models because it use the Open AI compatible.

Most services using OpenAI compatible protocols provide a models API. Please refer to https://platform.openai.com/docs/api-reference/models

Questions 2: Where are the server installed? I looked for them in my computer, but I can't find them. (Windows 10)

The Python Host mentioned above will be separate it out as planned

When we open a new conversation, if there is already something written in the message box and I click on tools, This shows the tools and it's correct, but the message is sent and it shouldn't do that.

I have successfully reproduced this error, and a fix will be available in next week's release. Thx

[onigetoc]

Questions 2: Where are the server installed? I looked for them in my computer, but I can't find them. (Windows 10)

The server is currently packaged within the application, but we have plans to separate it out."

Ok, because i'm wondering why are all these MCP servers separated? For example, on Cline, Roo Code, Google, Desktop ect. It should be possible to scan all these directories and find those already installed in our computer. (Why installing them multiple time?)

We are currently converting the MCP Host component from TypeScript to Python. At the time, the MCP Reload was designed to conveniently update everything at once during activation. After the MCP Host migration is complete, we will modify the MCP Loading mechanism to operate individually

Sorry, but I don't agree with that. It forces novice users to install Python, which adds complexity and another layer for novice users.

I'm working on this to help people adding and removing MCP servers.
https://preview--mcp-server-editor.lovable.app/

But for now i just found the Cline directory with all servers.
C:\Users\USERNAME\OneDrive\Documents\Cline\MCP

I should update Cursor if there's a existing folder.
i guess Claude Desktop are internal and we can not reach them too?

It far from finish but i will also add nodejs jsonrepaire, search filter ect,.
https://www.npmjs.com/package/jsonrepair
It should work adding multiple json server at once, work if you add

{
  "mcpServers": {

or not Remove from click.
Put the MPC servers by alphabetical order.

I'm writing now a MCP server to fix the Windows NPX issues, problems to make servers work on Windows.

Something needs to be done to make it easier because i think that non-technical users will not use mcp servers.

Thank your for all your help and work. :)

[kevinwatt]

Questions 2: Where are the server installed? I looked for them in my computer, but I can't find them. (Windows 10)

The server is currently packaged within the application, but we have plans to separate it out."

Ok, because i'm wondering why are all these MCP servers separated? For example, on Cline, Roo Code, Google, Desktop ect. It should be possible to scan all these directories and find those already installed in our computer. (Why installing them multiple time?)

I think there's a bit of a misunderstanding here. The server John mentioned is Dive's server. I guess he thought you were talking about our Front/Backend Server (aka, MCP Host+MCP Client).

But the Server you're referring to should be the MCP Server.

MCP Server is generally installed through package managers, such as uvx or npx. These user-level package managers don't install things repeatedly.

Actually, you don't have to install it separately. "uvx" and "npx" are integrated into your home dir. No matter if you run CludeDesktop or Dive, you are using the same MCP Server.

For example, Since I don't use Windows, on Linux, npx installs all the MCP Server in my $HOME/node_modules/ directory.

Also I don't know why cline repackage those package manager, and put all package under Cline\MCP. I don't use Cline.
I will talk with our team members, see if they know anything about it. :)

We are currently converting the MCP Host component from TypeScript to Python. At the time, the MCP Reload was designed to conveniently update everything at once during activation. After the MCP Host migration is complete, we will modify the MCP Loading mechanism to operate individually

Sorry, but I don't agree with that. It forces novice users to install Python, which adds complexity and another layer for novice users.

The user does not need to do anything. Python was included in the software package as early as the first version of Dive 0.1.0. You can't avoid TS/Py components in these complex SDK suites.

I'm writing now a MCP server to fix the Windows NPX issues, problems to make servers work on Windows.

Something needs to be done to make it easier because i think that non-technical users will not use mcp servers.

Good to know, and we'd be delighted to hear about any progress you make on this. Please feel free to share your developments with us anytime. It could potentially help simplify the installation process for the users.

[nicolaskrier]

Hello,

I am enjoying Dive AI as well and I have some comments and questions for you.

The user does not need to do anything. Python was included in the software package as early as the first version of Dive 0.1.0. You can't avoid TS/Py components in these complex SDK suites.

As a new user, I hadn't figured out this point yet. Could you update the documentation to clarify it?

I am currently taking my first steps with MCP and have configured a few MCP servers. I am using the Dive AI native application on an Apple MacBook Pro M4 with 48 GB RAM and Ollama running locally. I was considering using it even with all MCP servers disabled on Dive AI instead of Open WebUI, but I have found it really slow to respond: 12-13 seconds with Mistral Small 3 24B, while it took only 2-3 seconds using Open WebUI with the same model.
Have you experienced this kind of slowness while using Ollama locally?

Yesterday, I proposed a PR to improve the model selection interface while using Ollama locally. I wonder if it might be interesting to show only the model name and its tag if only one Ollama server is configured, especially if it is localhost. For LLM coming from HuggingFace, perhaps the HuggingFace short URL and the user repository do not need to be displayed.
What do you think about it?

Thank you for your time and for all the efforts you put into this project. Your work is greatly appreciated!

[johnfunmula]

As a new user, I hadn't figured out this point yet. Could you update the documentation to clarify it?

Currently, the Windows version of Dive includes Node.js, Python, the uvx environment, and executables. This means Windows users don't need to install these separately to use the MCP server driven by npx uvx. However, this is only available for Windows now, though other OS may have this feature in the future.

I am currently taking my first steps with MCP and have configured a few MCP servers. I am using the Dive AI native application on an Apple MacBook Pro M4 with 48 GB RAM and Ollama running locally. I was considering using it even with all MCP servers disabled on Dive AI instead of Open WebUI, but I have found it really slow to respond: 12-13 seconds with Mistral Small 3 24B, while it took only 2-3 seconds using Open WebUI with the same model. Have you experienced this kind of slowness while using Ollama locally?

We will investigate to find out where the problem is.

Yesterday, I proposed a PR to improve the model selection interface while using Ollama locally. I wonder if it might be interesting to show only the model name and its tag if only one Ollama server is configured, especially if it is localhost. For LLM coming from HuggingFace, perhaps the HuggingFace short URL and the user repository do not need to be displayed. What do you think about it?

merged.

[kevinwatt]

Hello,

As a new user, I hadn't figured out this point yet. Could you update the documentation to clarify it?

Here is the Backend and MCPHost. https://github.com/OpenAgentPlatform/Dive/tree/main/services

We probably won't spend too much time introducing this part at the moment. It's mainly because it's been redeveloped. In the future, this part may become an independent project. This will make the Dive backend (aka, Dived) run independently and provide the possibility of remote calls. It will have its own README and repository.

I am currently taking my first steps with MCP and have configured a few MCP servers. I am using the Dive AI native application on an Apple MacBook Pro M4 with 48 GB RAM and Ollama running locally. I was considering using it even with all MCP servers disabled on Dive AI instead of Open WebUI, but I have found it really slow to respond: 12-13 seconds with Mistral Small 3 24B, while it took only 2-3 seconds using Open WebUI with the same model. Have you experienced this kind of slowness while using Ollama locally?

Have you try ollama in console mode? like, ollama run modelname. does it feels slow too?
Dive has some system pre-loaded prompts. Similar to Claude Desktop. I'm not sure if this is the reason. But in your case it seems too slow. I just have a try to run ollama on my RTX 3060 and disable all my MCP servers., It looks fine with me.

https://youtu.be/JENX0k_CBOQ

Yesterday, I proposed a PR to improve the model selection interface while using Ollama locally. I wonder if it might be interesting to show only the model name and its tag if only one Ollama server is configured, especially if it is localhost. For LLM coming from HuggingFace, perhaps the HuggingFace short URL and the user repository do not need to be displayed. What do you think about it?

Thanks for the propose. We are open-minded about the model selector. If you have different ideas, please share them with us. At present, we still assume that users will switch between multiple models and multiple providers. If you have different design ideas, please feel free to share them with us.

[nicolaskrier]

In the future, this part may become an independent project. This will make the Dive backend (aka, Dived) run independently and provide the possibility of remote calls. It will have its own README and repository.

It would be highly beneficial to have two dedicated applications for the backend and frontend sides. This separation will enable users to implement their own backend application if desired. As a Java developer, I would appreciate having a frontend application with configurable MCP servers, allowing me to focus on developing my own backend application to experiment with Spring AI.

Have you try ollama in console mode? like, ollama run modelname. does it feels slow too?

I have tested with the Ollama desktop application started, Open WebUI, and then Dive AI as Ollama clients. I will investigate further when I have some spare time and share my findings with you.

Thanks for the propose. We are open-minded about the model selector. If you have different ideas, please share them with us. At present, we still assume that users will switch between multiple models and multiple providers. If you have different design ideas, please feel free to share them with us.

I agree with your perspective on the need for switchability across multiple models and providers. However, I believe it could be feasible to display only the relevant parts of the model based on the situation.
In my case, I have downloaded several LLM from Ollama directly or from Hugging Face and added them all to my local Ollama configuration. This might not be a common use case.
I plan to configure Mistral AI in the near future. For Mistral AI, I will need to disable some MCP servers for privacy reasons.
I believe having dedicated MCP servers configurations per model providers might be a nice feature.

[nicolaskrier]

I am currently taking my first steps with MCP and have configured a few MCP servers. I am using the Dive AI native application on an Apple MacBook Pro M4 with 48 GB RAM and Ollama running locally. I was considering using it even with all MCP servers disabled on Dive AI instead of Open WebUI, but I have found it really slow to respond: 12-13 seconds with Mistral Small 3 24B, while it took only 2-3 seconds using Open WebUI with the same model.
Have you experienced this kind of slowness while using Ollama locally?

I have created this dedicated issue for this topic. For Open WebUI, the results are worse than I thought, as I did not have the logs to provide insights.

[johnfunmula]

All the features mentioned in this issue have been completed in version 0.7.8