Open Deep Research

Open Deep Research is an open-source TypeScript project created by David Zhang (dzhng) and associated with Duet. It provides a minimal, readable implementation of a deep research agent — one that iteratively refines its research direction and dives deeper into a topic using LLMs, SERP queries, and web scraping. The repository is intentionally kept under 500 lines of code to remain easy to understand and build upon.

What It Is

Open Deep Research is a CLI-based autonomous research agent that takes a user query and two configurable parameters — breadth and depth — and recursively explores a topic until it has gathered enough information to produce a comprehensive markdown report. It belongs to the category of agentic research tools that chain LLM reasoning with live web search, as opposed to static knowledge retrieval. The core loop generates SERP queries, processes results into learnings and new research directions, and recurses until the depth parameter reaches zero.

How the Autonomy Loop Works

The agent follows a structured recursive flow:

• Input: User query + breadth (3–10, default 4) + depth (1–5, default 2)
• Follow-up questions: LLM generates clarifying questions before starting
• SERP query generation: Multiple targeted queries are generated per iteration
• Result processing: Extracts key learnings and new research directions from scraped content
• Recursive exploration: If depth > 0, the agent continues with new directions and accumulated context
• Report generation: All findings are compiled into a markdown report saved as report.md or answer.md

Concurrent processing is supported to run multiple searches and result extractions in parallel, with a configurable CONCURRENCY_LIMIT environment variable.

Model and API Flexibility

By default, the tool uses OpenAI's o3-mini model via the OpenAI API and Firecrawl for web search and content extraction. It also supports:

• DeepSeek R1 via Fireworks AI — automatically activated when a FIREWORKS_KEY is set
• Local LLMs — by pointing OPENAI_ENDPOINT to a local server (e.g., LM Studio) and setting OPENAI_MODEL
• Custom OpenAI-compatible endpoints — via OPENAI_ENDPOINT and CUSTOM_MODEL env vars, enabling OpenRouter, Gemini, and others
• Self-hosted Firecrawl — via FIRECRAWL_BASE_URL

Setup Path

The project runs in a Node.js environment and can also be deployed via Docker. Setup involves cloning the repository, running npm install, and configuring a .env.local file with API keys for Firecrawl and OpenAI (or alternatives). A Docker Compose workflow is also provided for containerized use.

Adoption Signal

According to GitHub data, the repository has accumulated over 19,000 stars and nearly 2,000 forks since its creation in February 2025, with active maintenance through April 2026. A community Python port is also available at a separate repository. The project is licensed under the MIT License, making it freely usable, modifiable, and distributable.