A full RAG stack wired into a single, chat-ready appliance

What it does

OpenRAG is a full-stack Retrieval-Augmented Generation platform built on FastAPI and Next.js. It lets users upload documents, process them, and query them through a chat interface backed by semantic search and large language models. Under the hood, it leans on Langflow for workflow orchestration, Docling for parsing messy real-world documents, and OpenSearch for the search index. The project also ships with Python and TypeScript SDKs, plus a built-in Model Context Protocol server so external AI assistants can query the knowledge base.

The interesting bit

The built-in MCP server is the stand-out detail: it exposes a streamable HTTP endpoint at /mcp for tools like Cursor or Claude Desktop, with no separate subprocess required. The README also admits the standalone openrag-mcp PyPI package is already deprecated in favor of this built-in route—a sign the integration strategy is still settling. Meanwhile, “agentic” workflows with re-ranking and multi-agent coordination are promised but not explained in any depth.

Key highlights

• Pre-packaged and ready to run out of the box, with core components already hooked together.
• Visual, drag-and-drop workflow builder powered by Langflow.
• Built-in MCP server over HTTP for direct assistant integration.
• Official async Python SDK and TypeScript/JavaScript SDK for embedding into other apps.
• Modular enterprise add-ons and OpenSearch-backed search scaling.

Caveats

• The README is high-level; specifics on how the “agentic” orchestration and “intelligent nudges” work are not explained.
• The MCP integration recently changed course—the standalone PyPI package is deprecated in favor of the built-in endpoint, so early adopters may have stale configs.

Verdict

Teams that want a batteries-included RAG interface with minimal integration work should look here. Pure infrastructure hackers who prefer to hand-tune their own vector stores and parsers will likely find the abstraction layer unnecessary.