mims-harvard/ToolUniverse

Democratizing AI scientists with ToolUniverse

What it solves

ToolUniverse provides a standardized way for Large Language Models (LLMs) to interact with scientific tools, enabling the creation of "AI scientists" that can perform complex data analysis, knowledge retrieval, and experimental design. It solves the problem of fragmented scientific software and APIs, integrating over 1,000 machine learning models, datasets, and scientific packages into a unified ecosystem.

How it works

The project uses an AI-Tool Interaction Protocol to standardize how LLMs identify and call tools. It supports a wide range of models (Claude, GPT, Gemini, Qwen, Deepseek) and integrates with the Model Context Protocol (MCP) for native server support. To manage the massive toolset, it features a "Compact Mode" that reduces the available tools to a few core discovery tools to save context window space. It also supports asynchronous operations for long-running scientific tasks like molecular simulations and a two-tier caching system (LRU and SQLite) for speed and reproducibility.

Who it’s for

It is designed for researchers and developers building AI agent systems for scientific discovery, particularly in fields like drug discovery, precision oncology, and bioinformatics.

Highlights

  • Massive Tool Library: Integrates 1,000+ ML models, datasets, and scientific packages.
  • MCP Integration: Native support for the Model Context Protocol.
  • Cachable & Async: Supports long-running tasks with progress tracking and two-tier result caching.
  • Pre-built Workflows: Includes 68 agent skills for specific research tasks like rare disease diagnosis.
  • Unified Literature Search: Combined access to PubMed, ArXiv, BioRxiv, and other scientific databases.