DeepWideResearch

Open Deep Wide Research

Agentic RAG for any scenarioCustomize sources, depth, and width

Why Do You Need Open Deep Wide Research?

In 2025, we observed 2 critical trends reshaping the Retrieval-Augmented Generation (RAG) tech stacks:

1. Traditional, Rigid, pipeline-driven RAG is giving way to more dynamic agentic RAG systems.

1. The emergence of MCP is dramatically lowering the complexity of developing enterprise level Agentic RAG.

However, a core pain point remains:

1. Developers still struggle to balance response quality, speed, and cost, as most agentic solutions offer a rigid, one-size-fits-all approach.

Based on these trends and the core pain point, the market needs a single, open-source RAG agent that is MCP-compatible and offers granular control over performance, scope, and cost.

We built Open Deep Wide Research to be that solution, providing one agent for all RAG scenarios. It gives you granular control over the core dimensions of agentic research:

• Sources: Connect custom data sources, from internal knowledge bases to specialized APIs.
• Deep: Controls response time and reasoning depth.
• Wide: Controls information breadth across your selected sources.

The "Deep × Wide" coordinate system also transparently predicts the cost of each response, giving you full budget control.

Example Scenarios:

User Story Settings Example Query Time Cost

Customer Service Bot Deep: ███░░░░░░░░░ 25%Wide: ███░░░░░░░░░ 25% "What glasses do you provide?" ~10s ~$0.01

Market Research Deep: ███░░░░░░░░░ 25%Wide: ████████████ 100% "100 Notion and Airtable alternatives" ~2-3min ~$0.10

Enterprise Analytics Deep: ████████████ 100%Wide: ████████████ 100% "What was the ROI of our latest marketing campaign?" ~5min ~$1.00

> If this mission resonates with you, please give us a star ⭐ and fork it! 🤞

Features

• Deep × Wide Control – Tune the depth of reasoning and breadth of information sources to perfectly match any RAG scenario, from quick chats to in-depth analysis.
• Predictable Cost Management – No more surprise bills. Cost is a transparent function of your Deep × Wide settings, giving you full control over your budget.
• MCP Protocol Native Support – Built on the Model Context Protocol for seamless integration with any compliant data source or tool, creating a truly extensible and future-proof agent.
• Self-Hosted for Maximum Privacy – Deploy on your own infrastructure to maintain absolute control over your data and meet the strictest security requirements.
• Hot‑Swappable Models – Plug in OpenAI, Claude, or your private LLM instantly.
• Customizable Search Engines – Integrate any search provider. Tavily and Exa supported out-of-the-box. As long as it supports MCP.

Get Started

Prerequisites

• Python 3.9+ and Node.js 18+
• API keys: Open Router (required), and Exa / Tavily (at least one)
• Recommended model: open-o4mini

Deployment Options

• API-only (Backend): If you only need the Deep Research backend as an API to embed in your codebase, deploy the backend only.
• Full stack (Frontend + Backend): If you want the full experience with the web UI, deploy both the backend and the frontend.

Backend

1. Copy the env template:

cp deep_wide_research/env.example deep_wide_research/.env

1. Edit the copied .env and set your keys:

# deep_wide_research/.env
OPENROUTER_API_KEY=your_key
# At least one of the following
EXA_API_KEY=your_exa_key
# or
TAVILY_API_KEY=your_tavily_key

> You can obtain the Tavily and Exa API keys from their official sites: Tavily and Exa.

1. Set up the environment:

cd deep_wide_research
python -m venv deep-wide-research
source deep-wide-research/bin/activate
pip install -r requirements.txt

1. Start the backend server:

python main.py

Frontend

1. Copy the env template:

cp chat_interface/env.example chat_interface/.env.local

1. Install dependencies and start the dev server:

cd chat_interface
npm install
npm run dev

1. Open the app:

Open http://localhost:3000 – Start researching in seconds.

Docker (Production)

docker-compose up -d

---

How We Compare

Feature OpenDeep WideResearch OpenAIDeep Research GeminiDeep Research ManusWide Research GenSparkDeep Research JinaDeepSearch LangChainOpen Deep Research

Depth × width controls D x W × × W × D ×

Open source ✅ × × × × ✅ ✅

MCP support ✅ ✅ × × × ✅ ×

SDK / API ✅ ✅ ✅ × × ✅ ✅

Local knowledge ✅ × × × × ✅ ✅

Model flexibility ✅ × × × × × ✅

Search engine flexibility ✅ × × × × × ×

Performance 5 5 4 3 4 4 3

Names are trademarks of their owners; descriptions are generalized and may change.

---

Deep Wide Research Archietecture

---

License

This project is licensed under the Apache License, Version 2.0. See the [LICENSE](LICENSE) file for details.

Copyright (c) 2025 PuppyAgent and contributors.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: puppyone-ai
• Source: puppyone-ai/DeepWideResearch
• License: Apache-2.0
• Homepage: https://www.deepwideresearch.com

Install and usage instructions live in the source repository linked above.