GPT Researcher is an open deep-research agent that produces a cited research report from a single query, drawing on the web, local documents, or both. It targets the gaps of manual research and single-LLM answering: stale training data, token limits that cut long reports short, and results skewed by a narrow set of sources.
The design follows a plan-and-solve pattern. A planner agent generates the research questions for a task, execution agents crawl sources for each question in parallel, and a publisher aggregates the summarized, source-tracked findings into the final report. A run gathers more than 20 sources and can exceed 2,000 words. A separate Deep Research mode explores a topic as a tree with configurable depth and breadth; a run takes about 5 minutes and costs about $0.40 with o3-mini at high reasoning effort.
You can use it as a Python package (Python 3.11 or later) through the GPTResearcher class, as a self-hosted server with a lightweight HTML frontend or a Next.js frontend, as a Docker Compose stack, or as a Claude Skill. The RETRIEVER setting can combine web search with MCP servers so that GitHub repositories, databases, and custom APIs become research sources. An OpenAI-compatible base URL lets you point it at local models or other providers.
Features
- Web and local research: reports draw on web crawling or on local PDF, text, CSV, Excel, Markdown, PowerPoint, and Word files under a DOC_PATH folder
- Deep Research: recursive, tree-shaped exploration with concurrent branches and shared context across them
- MCP retrieval: RETRIEVER=tavily,mcp runs hybrid web plus MCP research against configured MCP servers
- Report export: PDF, Word, Markdown, and other formats, with memory and context kept across the run
- Image handling: scraping and filtering of images for reports, plus optional inline illustrations generated with Gemini image models
- JavaScript-enabled scraping: pages that render client-side are crawled
- Multi-agent assistant: a LangGraph and AG2 variant, based on the STORM paper, in which a team of specialized agents produces a 5–6 page report in PDF, Docx, and Markdown
- Observability: LangSmith tracing for LangGraph runs and Monocle, an OpenTelemetry-based tracer, for end-to-end run records
- Assistant integrations: a Claude Skill and a separate MCP server (gptr-mcp) for Claude Desktop and other AI applications
