Harvard-backed ToolUniverse passes 500,000 AI agent analyses across 113 countries
ToolUniverse, an open science project from Harvard University, Harvard Medical School, and MIT, has completed more than 500,000 analyses by AI agents across 113 countries. The milestone reflects a shift in how researchers are deploying language models-moving beyond text generation toward systems that can verify findings, perform calculations, retrieve data, and run experiments using external tools.
Marinka Zitnik, Associate Professor at Harvard and Principal Investigator on the project, announced the update on LinkedIn. More than 236,000 of those analyses occurred in the last month alone.
What ToolUniverse does
The platform connects large language models-including Claude, GPT, Gemini, Qwen, and DeepSeek-to scientific tools, databases, and workflows. It uses a standardized protocol to help AI agents identify relevant tools, call them, interpret results, and chain them together into multi-step research workflows without requiring additional model training.
The system includes over 1,000 scientific tools covering literature search, drug discovery, precision oncology, rare disease diagnosis, molecular simulations, and pharmacovigilance. A case study showed an AI scientist moving from target identification to compound screening, property optimization, and patent assessment in hypercholesterolemia research-using databases like DrugBank, ChEMBL, and PubChem alongside machine learning tools.
Building infrastructure for AI scientists
ToolUniverse is one component of the broader Open AI Scientists initiative from Zitnik's lab. Related projects include Medea for multi-omics analysis and TxAgent for therapeutic reasoning-both built on ToolUniverse's tool ecosystem.
The platform includes tool discovery, tool composition, caching, asynchronous operations, and human-in-the-loop feedback. These safeguards address a practical challenge for universities and labs: autonomous research agents need reliable tools, structured workflows, and verification mechanisms to support scientific work beyond answering questions.
The project is available as open source with documentation and GitHub access. Its next phase will test how far open tool ecosystems can support reproducible, AI-assisted science as institutions experiment with agents capable of doing more than responding to queries.
For researchers interested in building these systems, AI for Science & Research courses cover the foundations of research automation and agent design.
Your membership also unlocks:
