// the find
melih-unsal/DemoGPT
🤖 Create agentic apps in a second with your prompts. Everything you need to create an LLM Agent - tools, prompts, frameworks, and models - all in one place.
DemoGPT bundles two different things under one repo: a Streamlit-app generator that turns a plain-English instruction into a LangChain-based app, and a separate agent/tool/RAG framework (demogpt_agenthub) with its own agents, tools, and vector store support. It's aimed at people who want to prototype an LLM app or agent without hand-writing LangChain boilerplate, not at production users.
The agenthub sub-library has a genuinely usable BaseTool/BaseAgent abstraction with a decent built-in tool set (search, weather, Python REPL, arxiv, YOLO vision) and RAG support across chroma, pinecone, and faiss. The ReactAgent's verbose reasoning trace, shown step by step in the README, is a real debugging aid for seeing why the agent picked a given tool. It has actual traction — 1.9k stars, pypi downloads, and citations in a few academic LLM-agent survey papers.
The core app-generation pipeline is dozens of hardcoded one-off prompt files per task type (hub_bash, hub_meteo, pal_chain, doc_load, etc.) — classic brittle 2023-era prompt-chaining, where adding a task means writing a new prompt module rather than composing existing ones. Test coverage is thin: only test_llms.py and test_rag.py exist, with no tests for the task-chain/code-generation pipeline, which is the most complex and failure-prone part of the project. The README still defaults examples to gpt-3.5-turbo and promises a Gorilla API integration 'within 2 weeks' that, per the to-do list, is still unchecked along with Llama2 support and the self-refining strategy — the roadmap has stalled well past what's advertised. It's unclear which of the two products (app generator vs. agenthub framework) is actually the maintained priority going forward.