An open-source GPT workshop with swappable brains and tools

What it does

OpenGPTs provides a self-hosted platform for building conversational agents akin to OpenAI’s GPTs and Assistants API. It wraps LangChain, LangGraph, LangServe, and LangSmith into a configurable stack where you choose the language model, vector database, tools, and retrieval algorithm rather than accepting OpenAI’s defaults. The result is a deployable backend and frontend for creating, testing, and sharing custom bots.

The interesting bit

The project treats “cognitive architecture” as user-serviceable parts. Instead of a single opaque agent, it ships three distinct LangGraph implementations—Assistant, RAGBot, and Chatbot—that you can edit directly in code. That makes it less a clone than a chassis: you get the GPT-like experience, but the wiring under the hood is exposed.

Key highlights

• Swap in any of 60+ LLMs, 60+ vector databases, or 100+ pre-built tools, or bring your own.
• Three built-in agent architectures: a flexible tool-calling Assistant, a retrieval-first RAGBot, and a Chatbot.
• Core feature parity with OpenAI’s GPTs for sandboxing, custom actions, knowledge files, drafts, and publishing via LangServe.
• Analytics and prompt debugging through LangSmith integration.
• Runs fully self-hosted with Postgres for persistence; Docker compose is supported.

Caveats

• The Assistant architecture only works reliably with a handful of top-tier models (GPT-3.5, GPT-4, Claude, Gemini) and can still “mess up.”
• A breaking checkpoint migration left pre-migration thread history visible in the database but inaccessible in the UI.
• Marketplace discovery and image-creation tools are listed as “coming soon.”

Verdict

Worth a look if you need a private, vendor-swappable alternative to OpenAI’s GPTs and are comfortable editing LangGraph code. If you just want a managed, polished no-code assistant builder, this is still closer to a framework demo than a finished product.