GPT Researcher is a deep research agent that takes a question, researches it across the web and local documents, and writes a long report with citations. It aggregates over 20 sources per run to reduce reliance on any single source, and every claim in the report traces back to a source.
The architecture splits work between a planner and execution agents. The planner generates a set of research questions, a crawler agent gathers information for each question in parallel, each resource is summarized and source-tracked, and a publisher filters and aggregates the summaries into the final report. A Deep Research mode explores a tree of subtopics with configurable depth and breadth, where findings seed new questions.
Before scraped passages reach the LLM, a context filter keeps only the passages that answer the question. Benchmarked options across 28 research tasks:
| Context filter | Relevant passages kept | Cost per report |
|---|---|---|
| Jev (default with a TypeSafe key) | 73% | $0.115 |
| Keyword BM25 (local fallback) | 51% | $0.116 |
| Embeddings | 46% | $0.117 |
| No filter | n/a | $0.192 |
It runs as a Python package, a server with a lightweight static frontend or a NextJS app, a Claude Skill, and an MCP server in a separate gptr-mcp repository. A multi-agent variant built with LangGraph and AG2 runs a team of agents from planning to publication.
Features
- Cited reports: research reports exceeding 2,000 words with source tracking, exported to PDF, Word, Markdown, and other formats
- Planner and execution agents: research questions are generated up front and crawled in parallel
- Deep Research: recursive, tree-like exploration of subtopics with configurable depth and breadth
- Local documents: researches PDF, plain text, CSV, Excel, Markdown, PowerPoint, and Word files, alone or combined with web sources
- Search engines: Tavily, Google, Bing, DuckDuckGo, Exa, arXiv, PubMed, and others
- MCP client: pulls from MCP servers such as GitHub repositories, databases, and custom APIs alongside web search
- Context filtering: Jev passage scoring by default, with keyword ranking or embeddings as alternatives
- Inline images: optional AI-generated illustrations in reports using Google Gemini image models
- Multi-agent assistant: a LangGraph and AG2 team that produces a 5–6 page report from planning to publication
- Observability: LangSmith tracing of LangGraph-based agent interactions
- Frontends: a lightweight HTML, CSS, and JavaScript frontend or a NextJS and Tailwind application with real-time progress
