Harvard-backed ToolUniverse passes 500,000 AI agent analyses across 113 countries
The open science project is building infrastructure for AI scientists, giving large language models access to scientific tools, databases, and research workflows.
ToolUniverse is an open science project designed to help AI agents access scientific tools, databases, and workflows
ToolUniverse, an open science academic research project involving Harvard University, Harvard Medical School, and the Massachusetts Institute of Technology, has passed more than 500,000 AI agent analyses across 113 countries.
Marinka Zitnik, Associate Professor at Harvard University and Principal Investigator on the ToolUniverse project, shared the update on LinkedIn, writing that more than 236,000 of those analyses took place in the last month alone.
The platform is designed to help researchers build "AI scientists" from large language models, including models such as Claude, GPT, Gemini, Qwen, DeepSeek, and open models. Its growth points to a wider shift in AI research infrastructure, where the question is moving beyond what a model can generate and toward what it can verify, calculate, retrieve, and test using external tools.
AI scientists move from chat to tool use
ToolUniverse describes itself as an ecosystem for creating AI scientist systems from any large language model. It uses an AI-Tool Interaction Protocol to standardize how AI agents identify and call tools, bringing together scientific models, datasets, APIs, and software packages for analysis, knowledge retrieval, and experimental design.
Zitnik wrote on LinkedIn: "ToolUniverse is going global."
She added that the platform began as "an open platform connecting AI agents to scientific tools, databases, and workflows" and is now becoming "a global, open foundation for AI agents."
The ToolUniverse site says the platform includes more than 1,000 scientific tools, covering areas such as literature search, drug discovery, precision oncology, rare disease diagnosis, pharmacovigilance, molecular simulations, and multi-step research workflows.
The academic paper "Democratizing AI scientists using ToolUniverse" describes the system as an infrastructure layer that wraps around a user-selected AI model, enabling it to identify relevant tools, call them, interpret outputs, and combine them into research workflows without additional model training.
Researchers build open infrastructure for scientific agents
The ToolUniverse project team includes researchers and affiliates from Harvard University, Harvard Medical School, the Broad Institute of MIT and Harvard, and MIT. Shanghua Gao is listed as Project Lead, with Zitnik named as Principal Investigator.
The wider Open AI Scientists initiative, from the Zitnik Lab at Harvard Medical School, frames AI scientists as autonomous systems that can reason, hypothesize, and experiment alongside human researchers. Its projects include Medea for omics analysis, TxAgent for therapeutic reasoning, ClawInstitute as an exchange for AI agents, and ToolUniverse as the foundation layer.
ToolUniverse is also connected to research on therapeutic reasoning. The platform lists TxAgent, which uses ToolUniverse’s scientific tools to solve complex therapeutic reasoning tasks, and Medea, which applies ToolUniverse tools to multi-omics analysis for therapeutic discovery.
In the ToolUniverse paper, researchers describe a case study in hypercholesterolemia where an AI scientist used the system to move from target identification to compound screening, property optimization, and patent assessment. The workflow used databases and machine learning tools including DrugBank, ChEMBL, Boltz-2, ADMET-AI, PubChem, and patent-mining tools.
Open AI research puts pressure on verification
For universities, research labs, and scientific computing teams, the ToolUniverse update highlights a practical challenge in AI adoption: research agents need access to reliable tools, structured workflows, and safeguards if they are going to support scientific work beyond text generation.
The project includes tool discovery, tool calling, tool composition, asynchronous operations, caching, Model Context Protocol integration, command-line access, and agent skills for predefined workflows. It also includes human-in-the-loop feedback and safety components, reflecting the risks of using autonomous systems in scientific and biomedical settings.
Zitnik wrote that researchers are using ToolUniverse "to build AI scientists, speed up analyses with agents, and explore new forms of scientific reasoning."
She added: "The future of science is bright."
ToolUniverse is available as an open science academic research project, with documentation, GitHub access, and setup instructions for AI agents. Its next phase will test how far open tool ecosystems can support reproducible AI-assisted science as universities and labs experiment with agents that do more than answer questions.
