UltraRAG is a low-code Python framework that builds complex retrieval-augmented generation pipelines by standardising RAG components as independent Model Context Protocol servers, aimed at researchers and teams doing industrial prototyping.

What it is

UltraRAG is the first lightweight RAG development framework based on the Model Context Protocol (MCP) architecture, jointly launched by THUNLP at Tsinghua University, NEUIR at Northeastern University, OpenBMB, and AI9stars. It is written in Python and released under the Apache-2.0 licence, and it sits in the AI and machine learning category as an AI development platform. The framework standardises core RAG components such as Retriever and Generation as independent MCP Servers, then pairs them with the workflow orchestration capabilities of the MCP Client. Version 3.0 was released on 23 January 2026, following 2.1 in November 2025 and 2.0 in August 2025; earlier code remains available on the v1 and v2 branches.

The concrete problem is glue code. Building a RAG pipeline with conditional branches and loops normally means writing and maintaining bespoke control logic. UltraRAG replaces that work with YAML configuration files, so developers describe sequential, loop, and conditional branch structures declaratively instead of hand-coding them. It also replaces the traditional chat interface with the UltraRAG UI, a visual RAG integrated development environment that combines orchestration, debugging, and demonstration.

Key capabilities

• Standardised MCP Servers for core RAG components including Retriever and Generation, coordinated by an MCP Client.
• YAML configuration for inference orchestration covering sequential, loop, and conditional branch control structures.
• UltraRAG UI Pipeline Builder with bidirectional real-time synchronisation between "Canvas Construction" and "Code Editing", plus online adjustment of pipeline parameters and prompts.
• Intelligent AI Assistant supporting pipeline structural design, parameter tuning, and prompt generation across the development lifecycle.
• One-click conversion of a constructed logic flow into an interactive dialogue system.
• Knowledge Base Management components for building custom knowledge bases for document Q&A.
• Multimodal support and a unified evaluation system introduced in version 2.1.

Who uses it and how

• Research groups exploring RAG ideas, a use case the project was designed around alongside industrial prototyping; it maintains a daily RAG paper digest updated every day.
• Teams building deep research pipelines locally, following the Lightweight DeepResearch Pipeline local setup tutorial published with a bilibili video and blog.
• Developers running writing agents on-device, using the open-sourced 8B AgentCPM-Report model released alongside the project.
• Industrial prototyping teams that want to concentrate on ideas with a few dozen lines of code rather than full pipeline implementation.
• Anyone standing up a document Q&A application from a custom knowledge base assembled in the UI.

Getting started

Installation is Python-based, and the project publishes a step-by-step installation and full RAG walkthrough video alongside a written guide, 00_Installing_and_Running_RAG.md. The facts provided do not name a package name, Docker image, or compose file, so the linked walkthrough is the authoritative setup path.

How it compares

No comparable products are named in the material provided, and no list of paid products it replaces is supplied, so it stands alone in this registry. The project does position itself specifically as the first lightweight RAG development framework built on MCP architecture.

When to use it — and when not to

The material does not state database, storage, or SMTP dependencies, nor does it state an exact package name or install command, so anyone evaluating it should read the linked installation guide before committing. Release cadence is fast, with four major versions between January 2025 and January 2026 and older code preserved on separate v1 and v2 branches, which means pipelines built against an older release must be pinned deliberately. Teams that want a hosted, managed service rather than a framework they install and operate themselves should look elsewhere, since no hosted option appears in the facts.

## Direct installation
pip install uv==0.12.0
## Download
curl -LsSf https://astral.sh/uv/0.12.0/install.sh | sh

git clone https://github.com/OpenBMB/UltraRAG.git --depth 1
cd UltraRAG

# Windows CMD
.venv\Scripts\activate.bat

# Windows Powershell
.venv\Scripts\Activate.ps1

# macOS / Linux
source .venv/bin/activate