What is co-scientist?

This is an open-source Python re-implementation of a research system described in a 2026 Nature paper by Google DeepMind called AI co-scientist. The original system uses multiple AI agents working together to take a research question written in plain language and produce a ranked list of novel scientific hypotheses. This repository follows the same agent structure and prompts described in the paper. The system has seven specialized agents, each with a distinct role. The Generation agent proposes new hypotheses by simulating a literature review and a scientific debate. The Reflection agent reviews each hypothesis for novelty and correctness. The Ranking agent runs an Elo tournament, the same scoring method used in chess, where hypotheses are compared against each other in simulated debates. The Evolution agent takes top-ranked hypotheses and tries to combine, simplify, or reframe them. The Proximity agent groups similar hypotheses together to reduce duplication and improve how the tournament pairs them up. The Meta-review agent synthesizes all feedback and writes the final research overview. A Supervisor coordinates all of these agents, manages a task queue backed by a local SQLite database, and writes the final output to a file. To run a session, you give it a research goal as a sentence, set a dollar budget and a time limit, and the system works through the agent cycle until it reaches one of its stopping conditions. The final output is a markdown file containing the research overview. There is also an optional web dashboard you can open in a browser that shows live progress. The tool works with many different AI providers including Anthropic, OpenAI, OpenRouter, Gemini, Groq, and others. You configure which provider to use and supply an API key. It also supports optional research tools like PubMed search, arXiv search, and web search if you provide the relevant API keys. The project is not affiliated with Google or the paper's authors. It is an independent implementation aimed at making the co-scientist approach available to researchers who want to run it themselves.