ToolUniverse: Democratizing AI Scientists
Build AI scientist systems from any LLM, integrating 1000+ scientific tools.
- Source repo
- mims-harvard/ToolUniverse
- Stars
- ★ 1.7k
- Last updated
- today
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 32/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformClaude Code · OpenAI API · Claude APIChatGPT (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Drug discovery researchers: Use ToolUniverse's tools and skills for virtual screening, drug interaction prediction, and other workflows to accelerate research.
- Not a fit if
- Users who want a GUI instead of terminal and code
- Those unwilling to configure MCP servers and API keys
- Source review
- 32/100 · Major gaps 8 safety controls not found
What does this agent do, and when should you use it?
ToolUniverse is an open-source ecosystem from Harvard's mims-harvard lab that turns any large language model (LLM) into an AI scientist. It standardizes how LLMs identify and call tools via the AI-Tool Interaction Protocol, integrating over 1000 machine learning models, datasets, APIs, and scientific packages for data analysis, knowledge retrieval, and experimental design. ToolUniverse supports universal AI models (Claude, GPT, Gemini, Qwen, Deepseek, and open models) and provides a native MCP (Model Context Protocol) server for seamless integration with various AI agents. It includes 68 pre-built research workflows (Agent Skills) for drug discovery, precision oncology, rare disease diagnosis, and more, plus a CLI (tu) and Python SDK for programmatic access. The project is Apache-2.0 licensed, with extensive documentation and community support (Slack, WeChat).
ToolUniverse provides an MCP server that exposes tools to AI agents, supporting asynchronous operations (e.g., protein docking, molecular simulations) with progress tracking, tool composition for sequential or parallel workflows, and a compact mode that reduces 1000+ tools to 4-5 core discovery tools, saving ~99% context window. It features two-tier result caching (in-memory LRU + SQLite) for 10x speedup and offline support. The CLI (tu) enables discovering, inspecting, running, and testing tools from the terminal. Unified literature search covers PubMed, Semantic Scholar, ArXiv, BioRxiv, and more. Agent Skills provide pre-built workflows for drug discovery and precision medicine. ToolUniverse also supports continuous expansion by registering new tools locally or remotely without extra configuration.
- Drug discovery researchers: Use ToolUniverse's tools and skills for virtual screening, drug interaction prediction, and other workflows to accelerate research.
- Bioinformatics scientists: Leverage its literature search and omics analysis tools for knowledge retrieval and multi-omics data mining to identify therapeutic targets.
- AI application developers: Integrate ToolUniverse into their own AI agents via MCP, avoiding the need to build tool-calling logic from scratch.
- Medical professionals: Use its skills for rare disease diagnosis or pharmacovigilance analysis to improve diagnostic accuracy and safety.
- Academic researchers: Use the Python SDK programmatically to build custom AI scientist systems for large-scale experiments and data analysis.
- Data science practitioners: Utilize its visualization, machine learning, and cheminformatics tools to handle complex scientific data and generate reports.
How do you install or deploy this agent?
For AI agent installation, open your AI agent (e.g., Claude Code) and run: Read https://aiscientist.tools/setup.md and set up ToolUniverse for me. The agent will guide you through MCP configuration, API keys, skill installation, and validation. Manual setup: Add the MCP server configuration to your MCP config file (using uvx with --refresh). Install agent skills via npx skills add mims-harvard/ToolUniverse. Claude Code users can install with: claude plugin marketplace add mims-harvard/ToolUniverse then claude plugin install tooluniverse@tooluniverse. For Python developers, install uv first, create a virtual environment with Python 3.12, then run uv pip install tooluniverse. For extra tools (ML, cheminformatics, visualization), install additional groups like [all] or specific ones like [ml].
How do you use this agent?
After installation, you can use ToolUniverse tools through your AI agent's MCP integration. For terminal use, the tu CLI allows you to discover tools (tu list), inspect details (tu show <tool>), and run them (tu run <tool>). In Python, import the SDK and create an AI scientist system, then call tools programmatically. For pre-built workflows, use the Agent Skills provided. For development, you can register new tools to extend functionality. Detailed usage is available in the official documentation, including CLI Reference and Python Developer Guide.
What are this agent's strengths and limitations?
- Integrates 1000+ scientific tools across domains, significantly lowering the barrier to building AI scientist systems.
- Model-agnostic, supporting Claude, GPT, Gemini, Qwen, Deepseek, and open models, with no lock-in.
- Native MCP server aligns with industry standards, facilitating integration with multiple AI agents.
- Provides 68 pre-built research workflows, ready for drug discovery and other applications.
- Advanced features like asynchronous operations, tool composition, and caching suit complex scientific tasks.
- Active community (Slack, WeChat) and continuous development by a reputable Harvard lab.
- Steeper learning curve for newcomers; requires understanding MCP configuration and tool-calling protocols.
- Base install only covers API and database tools; local ML, cheminformatics, and plotting require extra installs (e.g.,
[all]). - Some tools (e.g., async tasks) may depend on network and external APIs, potentially incurring costs.
- Documentation and community support, though rich, are specialized; scientific background may be needed to fully leverage.
How does this agent compare with similar options?
The repository does not explicitly mention competitors. However, based on functionality, alternatives like LangChain, AutoGen, and OpenAI function calling exist, but since they are not named, we do not elaborate further.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| ToolUniverse: Democratizing AI Scientists This agent | 32 · Major gaps | MCP serverFree + model costs | ★ 1.7k | today | Python | Claude Code · OpenAI API · Claude API |
| K-Dense BYOK | 49 · Major gaps | Desktop appFree + model costs | ★ 1.4k | 2d ago | TypeScript | ChatGPT · Codex · Claude.ai |
| Wisp Science | 71 · Some gaps | Desktop appFree + model costs | ★ 1k | today | Rust | Codex · Claude Code · OpenAI API · Claude API |
| Scientific Agent Skills | 51 · Major gaps | CLIFree + model costs | ★ 53k | 7d ago | Python | Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows: The repository does not provide explicit mechanisms for permission management, user confirmation, data flow transparency, sensitive data handling, dependency security, external effects, rollback, or source attribution. README mentions MCP configuration and API keys but does not specify least privilege or user confirmation. Dependency list is extensive but no security audit or vulnerability mitigation is mentioned. Therefore all trust criteria score 0.
Evidence shows: README and test files describe consistent tool registration and error handling mechanisms, but no actual runtime verification is provided. Dependency list is broad but no availability guarantees are stated. Test files include error handling tests but no user-visible error message examples are provided. Therefore self-consistency, dependency availability, and failure messages each score 1.
Evidence shows: README clearly targets AI scientists and developers, providing multiple usage scenarios (MCP, CLI, SDK). Capability boundaries are partially described but not all tool limitations are listed. Trigger precision is not detailed. Environment fit is good, supporting multiple Python versions and installation methods. Therefore audience and scenarios score 2, capability boundaries and trigger precision each score 1, environment fit scores 2.
Evidence shows: README provides clear information architecture with installation, usage, and documentation links. Installation notes are detailed, including manual and automatic methods. Naming is stable but no changelog is provided. Examples and FAQ sections exist but no known limitations are listed. License is Apache-2.0, version number exists, but no changelog is provided. Maintenance responsibility is clear but no update path is provided. Therefore information architecture, installation notes, examples and FAQ, and license each score 2, naming stability, known limitations, versioning, and maintenance responsibility each score 1.
Evidence shows: Output usability is not explicitly stated but CLI and SDK interfaces are provided. Marginal value is high, integrating 1000+ tools. Cost-benefit is not detailed but free open-source license is provided. Therefore output usability scores 1, marginal value scores 2, cost-benefit scores 1.
Evidence shows: Claims in README (e.g., 1000+ tools) are not supported by specific evidence. Test files exist but no independent verification is provided. Facts and inferences are not clearly separated. Therefore claim traceability, cross-source corroboration, and fact-inference separation each score 1.
- Not found in source: least-privilege scopingGrant only what the task needs: a dedicated account or read-only token, scoped to specific directories and repos.
- Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
- Not found in source: data-flow disclosureWatch which external services it contacts (proxy or firewall logs) and keep sensitive data out until you know where it goes.
- Not found in source: sensitive-data handlingUse dedicated, low-privilege, revocable API keys — never production credentials — and keep secrets out of logs.
- Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
- Not found in source: disclosed external effectsEstablish which external systems it writes to, sends to or changes, and verify with test accounts or repos before production.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- Not found in source: verifiable attributionInstall from the official repo or registry and check the publisher and URL to avoid look-alike packages.
- No permission management, user confirmation, or data flow transparency mechanisms are provided; assess security risks before use.
- Dependency list is extensive without security audit; potential known vulnerabilities.
- No rollback mechanism; difficult to recover after updates.
- No known limitations listed; users may misunderstand tool boundaries.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v1.5.7 Release v1.5.7
Release notes excerpt
What's Changed * Fix: an interrupted self-update broke the Desktop bundle's launch by @gasvn in https://github.com/mims-harvard/ToolUniverse/pull/759 * Add bounded PDB inventory and coincident-coordinate diagnostics by @gasvn in https://github.com/mims-harvard/ToolUniverse/pull/757 * Fix: the tool index lost its final newline on every build by @gasvn in https://github.com/mims-harvard/ToolUniverse/pull/760 * Close skill/tool gaps vs. the life-sciences tool marketplace and beyond, plus a new protocols.io tool by @gasvn in https://github.com/mims-harvard/ToolUniverse/pull/606 …View on GitHub ↗ -
v1.5.6 Release v1.5.6
Release notes excerpt
What's Changed * Call the account key a "connection key", as the website now does by @gasvn in https://github.com/mims-harvard/ToolUniverse/pull/751 * Say "tool run limit": the platform now counts only tool runs by @gasvn in https://github.com/mims-harvard/ToolUniverse/pull/752 * Fix: error results were cached and replayed on later calls by @Om-singhaI in https://github.com/mims-harvard/ToolUniverse/pull/754 * Fix custom PDB uploads for ProteinsPlus ProtoSS by @gasvn in https://github.com/mims-harvard/ToolUniverse/pull/749 …View on GitHub ↗ -
v1.5.5 Release v1.5.5
Release notes excerpt
What's Changed * Fix: republish the bundle when it installs an older release than this one by @gasvn in https://github.com/mims-harvard/ToolUniverse/pull/665 * Fix: let the bundle install on Python 3.14 by @gasvn in https://github.com/mims-harvard/ToolUniverse/pull/666 * Fix: correct the MCPB manifest licence and listing metadata by @gasvn in https://github.com/mims-harvard/ToolUniverse/pull/667 * Fix: validation errors named an array index instead of the parameter by @Om-singhaI in https://github.com/mims-harvard/ToolUniverse/pull/669 …View on GitHub ↗
FAQ
What are the prerequisites for installing and running ToolUniverse?
Does using ToolUniverse incur any costs?
How can I integrate ToolUniverse with my custom AI agent?
What should I do if a tool fails or returns inaccurate results?
tu CLI to test and debug tools. Check the caching logs for issues. Additionally, community support via Slack and GitHub Issues can help resolve problems.