deep-research

Why we included this project

Have you ever watched an LLM answer confidently from a single prompt and wished it would actually go read the sources? This project is a compact demonstration of how to build that behavior. It runs a recursive loop: generate search queries from your question, scrape and process the results into learnings, then decide whether to go deeper or compile everything into a markdown report. The whole agent is deliberately kept under 500 lines of code, which makes it unusually easy to read end-to-end and adapt to your own stack. Developers who want a reference implementation of an iterative research agent, or a starting point to bolt onto their own tooling, will find the depth and breadth parameters easy to tune and the code cleanly split into stages that are simple to follow. It is a learning-oriented codebase rather than a hosted service, so expect to bring your own Firecrawl and OpenAI-compatible API keys.