open-coscientist 0.2.0

Open LangGraph-based framework for multi-agent research hypothesis generation, adapted from Google Research's AI Co-Scientist.

Open Coscientist

AI-powered research hypothesis generation using LangGraph

Open Coscientist is an open adaptation based on Google Research's AI Co-Scientist research paper. This project provides an implementation that generates, reviews, ranks, and evolves research hypotheses using the multi-agent architecture described. It orchestrates 8-10 specialized AI agents through a LangGraph workflow and aims to produce novel hypotheses grounded in scientific literature.

Demo

In this demo we use Open Coscientist to generate hypotheses for novel approaches to early detection of Alzheimer's disease. Click to watch the full demo on YouTube.

Standalone operation

The engine works with any LLM and can run without external data sources.

For high-quality hypothesis generation, the system provides an MCP server integration to perform literature-aware reasoning over published research. See MCP Integration for setup and configuration details, and to run the basic reference MCP server.

Quick Start

Installation

pip install open-coscientist

Set your API key (any LiteLLM-supported provider):

export GEMINI_API_KEY="your-key-here"
# or: export ANTHROPIC_API_KEY="your-key-here"
# or: export OPENAI_API_KEY="your-key-here"

For development, see CONTRIBUTING.md.

Note: for the any literature review to run, you must provide an MCP server with literature review tools/capabilities. You can use the provided reference implementation MCP Server. Otherwise, no published research will be used.

Model Support: Uses LiteLLM for 100+ LLM providers (OpenAI, Anthropic, Google, Azure, AWS Bedrock, Cohere, etc.). May need to tweak some constants.py token usage and other params, such as initial hypotheses count, in order to work with less powerful models.

Basic Usage

import asyncio
from open_coscientist import HypothesisGenerator

async def main():
    generator = HypothesisGenerator(
        model_name="gemini/gemini-2.5-flash",  # default model if not provided
        max_iterations=1,
        initial_hypotheses_count=5,
        evolution_max_count=3
    )

    async for node_name, state in generator.generate_hypotheses(
        research_goal="Your research question",
        stream=True
    ):
        print(f"Completed: {node_name}")
        if node_name == "generate":
            print(f"Generated {len(state['hypotheses'])} hypotheses")

if __name__ == "__main__":
    asyncio.run(main())

See examples/run.py for a full example cli script with a built-in Console Reporter. Remember, you must run the literature review MCP server for any literature review to be included in the hypothesis generation.

Features

• Multi-agent workflow: Supervisor, Generator, Reviewer, Ranker, Tournament Judge, Meta-Reviewer, Evolution, Proximity Deduplication
• Rich hypothesis output: Each hypothesis includes text, explanation (layman summary), literature_grounding with structured [C*] citations, and experiment (suggested validation design)
• Literature review integration: Optional MCP server provides access to real published research; structured citations resolve to full source metadata
• Domain-agnostic customization: YAML-based configuration to bring your own MCP servers, literature sources, and domain-specific prompt guidance — no code changes needed (see Domain Customization)
• Real-time streaming: Stream results as they're generated
• Intelligent caching: Faster development iteration with LLM response caching
• Elo-based tournament: Pairwise hypothesis comparison with Elo ratings
• Iterative refinement: Evolves top hypotheses while preserving diversity
• Post-generation enrichments: Attach domain-specific data (e.g., related CVEs, knowledge graph statements) to each hypothesis via configurable tool calls

The workflow automatically detects MCP availability and adjusts accordingly. Functional reference MCP server included in mcp_server/ directory.

Documentation

• Architecture - Workflow diagram, node descriptions, state management
• MCP Integration - Literature review setup and configuration
• Generation Modes - Three generate node modes explained, and parameters to enable them
• Configuration - All parameters, caching, performance tuning
• Domain Customization - Adapting to new domains (cybersecurity, bioinformatics, etc.) via YAML config
• Literature Review Tools Configuration - YAML schema reference for custom MCP servers and multi-source literature review
• Logging - File logging, rotating logs, log levels
• Development - Contributing, node structure, testing

Node Descriptions

Node	Purpose	Key Operations
Supervisor	Research planning	Analyzes research goal, identifies key areas, creates workflow strategy
Literature Review (Recommended)	Academic literature search	Queries databases (PubMed, Google Scholar), retrieves and analyzes real published papers (requires MCP server; without it, uses only LLM's latent knowledge)
Generate	Hypothesis creation	Generates N initial hypotheses using LLM with high temperature for diversity
Reflection (Recommended)	Literature comparison	Analyzes hypotheses against literature review findings, identifies novel contributions and validates against real research (requires literature review)
Review	Adaptive evaluation	Reviews hypotheses across 6 criteria using adaptive strategy (comparative batch for ≤5, parallel for >5)
Rank	Holistic ranking	LLM ranks all hypotheses considering composite scores and review feedback
Tournament	Pairwise comparison	Runs Elo tournament with random pairwise matchups, updates ratings
Meta-Review	Insight synthesis	Analyzes all reviews to identify common strengths, weaknesses, and strategic directions
Evolve	Hypothesis refinement	Refines top-k hypotheses with context awareness to preserve diversity
Proximity	Deduplication	Clusters similar hypotheses and removes high-similarity duplicates

Literature Review and Domain Customization

The bundled MCP server provides a PubMed reference implementation. The system is domain-agnostic: a YAML configuration file controls which MCP servers, literature sources, and prompt guidance are used — no code changes needed. Example configurations are included for biomedical (INDRA + PubMed), cybersecurity (arXiv + Google Scholar + NVD), and multi-source academic research.

See MCP Integration to set up literature review, and Domain Customization to adapt to your research area.

Attribution

Open Coscientist is a source-available implementation inspired by Google Research's AI Co-Scientist. While Google's original system is closed-source, this project adapts their multi-agent hypothesis generation architecture from their published research paper.

Reference:

• Blog: Accelerating scientific breakthroughs with an AI Co-Scientist
• Paper: Towards an AI co-scientist

This version provides a LangGraph-based implementation. It includes some optimizations for parallel execution, streaming support, and caching.

Citation

If you use this work, please cite both this implementation and the original Google Research paper:

@article{coscientist2025,
  title={Towards an AI co-scientist},
  author={Google Research Team},
  journal={arXiv preprint arXiv:2502.18864},
  year={2025},
  url={https://arxiv.org/abs/2502.18864}
}