No issues but suggestions

[onigetoc]

1. It would be interesting to create workspaces to work with folders and files on the computer. Having a dedicated place to work, so being able to create workspaces.

2. It would also be useful to keep the tools that have been used and always retain them in memory along with their results because if we go back in history, we lose them.

3. And I don’t know if you called them Artifacts or Canvas, but like Claude and OpenAI, we should be able to force the opening of the Canvas or Artifacts (or whatever name you gave it) to display the results we want. For example, if I ask,

• "Create an HTML page on this topic and show me the result," something like that.
• make me a markdown on this.. an show me the result.
• i want to make a trip to Tokyo and create me a 7 days trip with activité to do and show me the result in a html file. (It will show the result in the right window.

Also it will be more interesting to see the preview by defaullt instead of the code by default like it do now. Or may be smart coice by the AI to show code or preview according to the subject. I.E. Showing code first for coding questions.
4. Accepting user context general prompt (textarea)
5. Do this client accept the3 functions? Tools (obvious already working one), Resources, Prompt.

QUESTION:
Is the JSON of the tools accessible externally? I want to create a tool installer and a tool repairer, so i should have access to the JSON MPC Server json file to modify them, just like I can right now do with CLAUDE, CLINE, ROO CODE, Cursor, Windsurf ect. for example.

[kevinwatt]

Yes, many of these suggestions are already on our radar and in active development.

1. Workspaces/Projects: We are planning workspace or project functionality based on a directory structure. Each category will contain multiple Tasks, Documents (images, videos, PDFs, docs, etc.), and Manuals - primarily designed to integrate with our future Scheduling system. This feature is currently positioned in the later stages of our RoadMap.

2. Tool Memory: Version 0.8.0 will introduce Host changes where all tool responses will be recorded. MCP results within discussion sessions will be saved as RAG data. This ensures that even when continuing discussions or resuming historical sessions, previous function calling results remain accessible.

   • In pre-0.8.0 versions, special MCPs or handling multiple PDFs can easily overflow the content windows, which is why we've prioritized this improvement for immediate implementation.

3. Artifacts/Canvas: We don't have an official name for this sidebar yet - during development we simply call it the "right window." The issue where code displays instead of previews is a known bug (some versions do automatically switch to preview mode). I think @johnfunmula is responsible for this feature will address this issue soon.

4. Custom Instructions: We already have a feature similar to ChatGPT's custom instructions in our LLM settings page. Is this what you're looking for? If so, it might be a matter of unclear naming or placement - we'll have our UX review this to improve discoverability.

5. Functions Support: Regarding Resources and Prompt functions - could you specify what direction you're envisioning? For example:
   • Resources: Are you thinking of a shared file library with RAG capabilities?
   • Prompt: Are you looking for a prompt library with version control?

Regarding External JSON Access (Your Question):
In version 0.8.1 or 0.8.2, we'll introduce a CLI version of Dive that serves as a unit testing tool for the backend. This will allow control of all UI-manipulable values through command options.

Additionally, Dive's frontend will support commands like "claude mcp" in the console for adding, deleting, or enabling/disabling MCP tools. This should provide the external tool management capabilities

[onigetoc]

All about Tools, Ressources, Prompts for MCP client and server.


https://huggingface.co/blog/Kseniase/mcp

You probably know the majority of these stuffs, but also some information, more or less, on what I was talking about.
https://www.devshorts.in/p/how-to-build-your-own-mcp-server

About memory or chat history,
Sometimes, let's say I use the Brave mcp server to search with Brave. Sometimes the LLM will show me the links and sometimes not (It can also depend on which LLM we used, of course.). It seems that the result of the server is not in memory, only the text the LLM provided. Let's say that the LLM does not show me the links. I can not ask for an info on this link (to fetch the URL fro example) is in the result of the server, but not in the text the LLM provided. It would seem that he is not aware of these links because the LLM did not mention these link in the respond for this example.
Or another example, sometimes I do a search on YouTube with a MCP server for YouTube, but the LLM gave me only the titles and if I ask for information on the description or links to go deeper, it will not be aware of these infos, even if it was in the final result of the MCP server. I may be wrong, but since the LLM gave only a part of the info, he does not seem to know everything it got about this search in a following conversation.
Because, for example, I have a YouTube MCP server downloader. If he doesn't display the links in the conversation, he won't know what to download at all because, for example he only gave me the titles and the description without the links. If I ask him to download the second video, he will ask me for the link because he is not aware of the links even if they are in the json result response.
But at the same time, it would greatly increase the context and the tokens usage. So I don't really know what's best.

If i test it with Cline or Roo Code, they seem to have all the context receive from the json response of the mcp server even if it gave me a short response in the conversation.

[chymian]

hi there,
I just found out about this fantastic prg. and this FR addresses the most of my wishes.

one thing is a system wide memory, like letta included MemGPT to remember general stuff about a user's preferences. - next to workspace specific -memory/rag.
i.e. don't need to tell every (Recipe-) workspace (group) the same static parameters, like that I'm vegetarian and the stuff I'm allergic to.
a possibility to commit an information/answer into general memory (group) and an additional feature on top for the ability to turn these personalized groups of general memory off for specific groups of workspace, to make it act generic again would IMHO bring it closer to be a real world Assistant, a Personal AI Assistant, which can grow/learn and adapt over time (like Jarvis) – if that's one of your your goals.

[kevinwatt]

Hi everyone,

Thank you for your patience. We have successfully completed the PyHost transition. This task proved more challenging than initially anticipated, which required additional time to implement properly.

With the current version, we are beginning our research into RAG and Letta integration.

One issue we've identified concerns some popular local models (not Claude/GPT/Gemini) available on Ollama or OpenRouter, such as Qwen3 and DeepSeek. These models have limitations with tool calling support, which presents challenges since Letta's functionality relies heavily on tool calling capabilities.

We're considering adding a memory capability indicator in the input textarea that would show users which models fully support Letta's assistance features. This would help set appropriate expectations about memory-enhanced capabilities depending on the model in use.

These are our initial thoughts on addressing these integration challenges. We welcome any feedback on this approach.

[chymian]

talking about memory and RAG. just learnd about https://docs.morphik.ai/introduction:

What is Morphik?

An overview of Morphik

Morphik is a database that makes it easy to create fast, versatile, and production-ready AI apps and agents.

Key features include:

• First class support for Unstructured Data: Unlike traditional databases, Morphik allows users to directly ingest unstructured data of all forms - including (but not limited to) videos and PDFs. We’ve built research-driven custom algorithms to ensure state-of-the-art retrieval accuracy.
• Persistent KV-caching: For documents you reference often, you can process them once and freeze the LLM’s internal state such that you can use it again later. This helps drastically reduce compute costs as well speed up model responses.
• Out of the box MCP support: Morphik has built in support for the open-source Model Context Protocol - so you can integrate your knowledge with any MCP client in a single click.
• Natural Language Rules Engine: Define natural language rules that dictate how your data is ingested and queried. Think of it like defining schemas for your unstructured data.
• User and Folder Scoping: Organize and isolate your data with multi-user and folder-based access controls. Create logical boundaries for different projects or user groups while maintaining a unified database.
• Completely Open-source: You can check out Morphik core code here!.
• Flexible Model Registry: Easily register and use hundreds of different AI models across your application with our new registered models approach. Mix and match models by task (e.g., use smaller models for simpler tasks, powerful models for complex reasoning) through a simple configuration.