mims-harvard/ToolUniverse

ToolUniverse – A Platform for Building AI‑Scientist Agents

What it is – ToolUniverse is an open‑source ecosystem that lets you turn any large language model (LLM) into an AI scientist. It does this by providing a standardized AI‑Tool Interaction Protocol (AITIP) and a catalog of more than 1 000 ready‑to‑use scientific tools – ML models, cheminformatics packages, literature‑search APIs, and other data‑analysis utilities. The platform ships a CLI (tu) and a Python SDK so developers can discover, compose, and run tools from scripts or from within an LLM‑powered agent.

Key components

• MCP (Model Context Protocol) server – a lightweight local server that mediates tool calls from the LLM, handling authentication, transport, and result caching.
• Tool catalog – each entry describes the tool’s inputs, outputs, and a JSON schema that the LLM can fill in. Tools cover domains such as drug discovery, omics analysis, literature search, and general‑purpose ML.
• Async execution & composition – long‑running tasks (e.g., protein docking) run in the background with progress updates; multiple tools can be chained or run in parallel to build complex workflows.
• Compact mode – automatically reduces the catalog to a handful of core discovery tools, saving the LLM’s context window.
• Two‑tier caching – in‑memory LRU cache plus a persistent SQLite store, keyed by a fingerprint of each tool’s arguments, giving ~10× speed‑ups and offline reproducibility.

How you use it

1. Install – recommended via an AI agent (Claude, etc.) or manually with the uv package manager (uv pip install tooluniverse). Optional extras ([ml], [visualization], …) add domain‑specific dependencies.
2. Configure – add a small entry to your MCP config file so the local server knows how to launch the tooluniverse command.
3. Run tools – from the terminal with tu <subcommand> (discover, inspect, run, test) or from Python with the SDK (from tooluniverse import ToolClient).
4. Build agents – attach the MCP server to any LLM (Claude, GPT, Gemini, Qwen, Deepseek, open models). The LLM learns the protocol, decides which tool to call, sends a JSON request, and receives the result.

Notable projects built on top of ToolUniverse

• TxAgent – a therapeutic‑reasoning agent that uses the catalog to design and evaluate drug candidates.
• Medea – an omics‑focused AI agent for multi‑omics analysis and therapeutic target discovery.

Who’s behind it – Developed at Harvard Medical School’s Zitnik Lab, led by Shanghua Gao and Marinka Zitnik, with contributions from several graduate students and post‑docs.

Documentation & community – Full docs, API reference, and a step‑by‑step guide are hosted at zitniklab.hms.harvard.edu/ToolUniverse. A Slack workspace, WeChat group, and public GitHub issues provide support. The project is listed in the Model Context Protocol registry and has a PyPI package (tooluniverse).

Why it matters – By abstracting away the plumbing needed for an LLM to call scientific software, ToolUniverse lowers the barrier for researchers to create AI‑augmented scientists that can retrieve literature, run simulations, and analyse data autonomously, accelerating hypothesis generation and experimental planning.