Hi, OpenAI Community, I really don’t know where else to post this. I am not “involved” in AI, but have built something that I am wondering if anyone else will find useful, or alternatively, something better exists that I can use for myself. It is an attempt to create a platform for extensible agentic applications (I call them Components) under a unified umbrella of MCP and persistence.
This started with my frustration in trying to create actually useful agents using chatGPT. I noticed that the worst thing about most agents I’ve tried to make is that they are awful, and by awful I mean psychotic and non-deterministic. Mine lied to me for weeks about having some secret structured data storage, which held things like log data and other tabular stuff, and justified it with all kinds of hidden infrastructure that *does not exist*. That hallucination, and the realization that there is no “native” structured data storage (or vector memory!) from the foundation models’ demo apps, started me on this Project.
For context on the level of sophistication and “market maturity” of the facilities I’ve built (such as they are), I have about 3 decades of experience in IT, with a heavy chunk of data processing but little to no data science… Mostly line-of-business app dev and data monkey work in the legal and healthcare industries. I made this mostly as a toy and a demo project to help me learn “AI stuff”. But I think it might really be useful to others, or at least, I would like to know what competing products exist that do similar things.
Features:
-Terminal command prompt primary interface which implements dynamic menus for direct API commands and passes prose commands to the LLM for interpretation and further calls;
-Persistent, deterministic, rich and dynamic control dashboards (jsx snippets are kept in the DB and inserted at request time by the OS);
-Postgres DB with CRUD API (you define the exposed tables static for now; obviously Maple should be able to create her own tables);
-Side and bottom panels to display different dashboards, to give the platform some structure and coherence;
-The beginnings of Prompt versioning and feature-driven prompt injection;
-Vector memory facility (using pgvector) with standard CRUD operations plus semantic search;
-Langchain / ReAct loop with dedicated debug screen for deep introspection of MCP calls (maybe at runtime but definitely as a snapshot replay);
-Parallelizable worker units and a task queue for doing work;
-A bunch of other cost and processing metrics I’m probably forgetting about.
My “demo app”, the use case I built this system to implement, is essentially a document repository Component, and a museum of sorts: 18 million message board posts from The Motley Fool collected between 1997 - 2010, something that has been on my mind to expose to the public since that website took their old boards offline (breaking my heart).
The original announcement is here: Update and Introduction - Macro Economic Trends and Risks - Motley Fool Community , and there is a better description plus a few screenshots of my TMF and worker dashboards.
A live instance of it is running at https://whafa.com . But I wouldn’t exactly call it production-ready. Also I currently only support Google Oauth, sorry. But if you want to check it out, try signing in (you can’t do much unauthenticated) and then typing “mdebug -s” to open the ReAct debugger, then ask Maple a question about… I dunno. Financial planning. Index funds. Most of the data is yet to be embedded, but I also have all the original DocText indexed.
So… Do you have any interest in using this? What about working on it? What about just looking at it and telling me I’ve reinvented the wheel and all this functionality and more is available elsewhere. The code is in a Private repo on GitHub, and it’s my first attempt at a a collaborative coding anything, so I am really apprehensive about making it public. But I could open it for a few people who might be looking for this kind of thing, and see how that goes? Anyway, thank you for reading this far, or if you used a bot to summarize it, thank you to the bot.
Thank you, and please be nice,
-Richard
P.S. obviously I never would have done any of this without the amazing coding tools released in just the past few months by OpenAI (et al). If you worked on these tools, THANK you. You have changed the world.