open_deep_research: Build Configurable Deep Research Agents

Overview

open_deep_research is a configurable agent for turning a research question into a synthesized report. It coordinates search, summarization, compression, and report writing, addressing the work of gathering and organizing information across multiple sources.

It is aimed at developers building or evaluating research agents, rather than people looking for a standalone research service. Model providers, search APIs, and MCP servers can be configured, and the project can be run locally through LangGraph Studio or deployed on LangGraph Platform.

Key Features

• Uses separate configurable models for search research, result summarization, compression, and final report writing.
• Supports chat models through LangChain's init_chat_model() API.
• Offers configurable search tools, with Tavily as the documented default.
• Supports MCP servers and native web search for Anthropic and OpenAI.
• Provides a LangGraph server and Studio workflow for local testing and configuration.
• Includes evaluation scripts for running and exporting results for Deep Research Bench.
• Retains legacy workflow and multi-agent implementations for comparison.

Use Cases

• Developers prototyping a research assistant who need to change models or search providers without rebuilding the workflow.
• Teams integrating MCP tools or specialized search APIs into a research process.
• Researchers comparing model configurations on multi-step research and report-generation tasks.
• Organizations testing a configurable research agent locally before deploying it through LangGraph Platform or Open Agent Platform.

Getting Started

See the README for setup and configuration details. The documented local quickstart is:

git clone https://github.com/langchain-ai/open_deep_research.git
cd open_deep_research
uv sync

Considerations

• You need credentials for the chosen language model and search services. The documented default search provider is Tavily.
• Selected models need to support structured outputs and tool calling.
• Running the full 100-task benchmark can cost approximately $20-$100, depending on model selection, according to the README.
• The repository is marked archived in the supplied repository data. Check its current status and dependencies before relying on ongoing maintenance.
• The README describes earlier implementations as less performant than the current implementation; treat the legacy code as reference material.