M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
T

ToolUniverse MCP Server

Official
2,500+ scientific tools for AI scientists.
How to install ↓ Compare GitHub source repository ↗
Category
Other #142 of 230
Stars
★ 1.7k Very popular
Transport
stdio (local process)
Runtime
Python 3.12+ · Prebuilt binary uv (uvx launcher)
Credentials
Optional API key
License
Apache-2.0
Last commit
51FMRS · D

ToolUniverse MCP Server targets science-focused AI agents, with broad coverage and support for tool discovery, execution, composition, asynchronous operations, and scientific data sources. It is best suited to research and development teams able to configure APIs and validate results.

Strongest · Documentation 13/20 Weakest · Reliability 7/20

Reliability
7/20
Security and permissions
8/20
Maintenance
11/20
Documentation
13/20
Setup experience
12/20
Why each score
Reliability 7/20
The manifest provides a PyPI package, uvx runtime hint, stdio transport, and compact-mode arguments, while the README describes a plausible startup and tool-discovery path. However, the supplied material contains no committed CI, tests, or MCP handshake evidence; it also inconsistently describes the tool count as “2,500+” and “1000+.” Error handling, dependency failures, and actual declared-tool behavior are unverified, so points are deducted.
Security and permissions 8/20
The manifest marks API keys as optional secrets and discloses dependencies on multiple external LLM, database, and scientific APIs. There is no evidence here of real tokens, credential theft, or overtly malicious behavior. However, the broad tool surface implies extensive external-network and potentially computational or research activity, while least privilege, data boundaries, outbound-data disclosure, dangerous-operation confirmation, and key isolation are not evidenced. The score is reduced for these missing controls.
Maintenance 11/20
The project states an Apache-2.0 license and supplies version 1.4.1, a GitHub repository, maintainer, leaders, contributors, GitHub Issues, and community channels, establishing basic ownership and update paths. The material does not provide commit history, release cadence, dependency-update practice, issue-response evidence, vulnerability disclosure, or a security-response process. The 26 open issues do not by themselves demonstrate maintenance quality, so points are deducted.
Documentation 13/20
The README supplies installation methods, an MCP configuration example, Python SDK and CLI references, documentation links, category information, feature descriptions, and an API-key inventory, giving it broad coverage. It lacks auditable tool parameters and return schemas, systematic cost and rate-limit disclosure, platform compatibility details, permission and data-flow boundaries, and troubleshooting evidence. Its inconsistent tool-count claims also reduce confidence, so points are deducted.
Setup experience 12/20
The project offers a uvx/stdio configuration, a Python installation path, client examples, category-selection arguments, and agent-assisted setup, making the basic mainstream MCP connection reasonably clear. Practical use may still require installing uv, selecting categories, configuring optional service credentials, and handling additional dependency groups. Cross-platform validation, connection-failure diagnostics, and reproducible end-to-end execution evidence are absent, so points are deducted.

Static review · not runListed 2026-08-14

Read the FMRS scoring method →

Fit and risk

What it can accessUses the network

Best for

  • Researchers needing unified access to many scientific APIs, databases, models, and packages
  • Developers building life-science, precision-oncology, or drug-discovery agents
  • AI scientist projects using the CLI, Python SDK, or MCP integration

Not for

  • Users who need only a few general-purpose office or development tools
  • Workflows that cannot configure any external scientific-service credentials
  • Scenarios where AI-generated results will be treated as validated scientific conclusions

Required permissions

  • Local execution of uvx and the ToolUniverse Python package
  • Access to external APIs, databases, or model services used by selected tools
  • Optional API keys for model, literature, patent, biomedical, and structure-prediction services
  • Some tools may require locally installed extras or additional dependencies

Risks and side effects

  • Scientific retrieval, analysis, and reasoning outputs require human validation
  • Configured API keys are used with their corresponding external services
  • External services may impose rate limits, credential requirements, or availability changes
  • Loading many tools or running long tasks may increase resource usage and context-window pressure

Setup

Before you start

Runtime：Python 3.12+ · Prebuilt binary uv (uvx launcher)

OPENAI_API_KEY optionalsecret OpenAI API key for LLM-based tool finding and embedding features; optional, from platform.openai.com.
ANTHROPIC_API_KEY optionalsecret Anthropic API key for Claude-based features; optional, from console.anthropic.com.
GEMINI_API_KEY optionalsecret Google Gemini API key for Gemini-based features; optional, from Google AI Studio.
OPENROUTER_API_KEY optionalsecret OpenRouter key for accessing 100+ LLM models; optional, from openrouter.ai.
AZURE_OPENAI_API_KEY optionalsecret Azure OpenAI key (alternative to OpenAI); optional, from Azure Portal.
HF_TOKEN optionalsecret HuggingFace token for accessing models and datasets; optional, from huggingface.co/settings/tokens.
NCBI_API_KEY optionalsecret NCBI/PubMed key raising rate limits from 3 to 10 req/sec; optional, register at ncbi.nlm.nih.gov/account/.
SEMANTIC_SCHOLAR_API_KEY optionalsecret Semantic Scholar key raising rate limits from 1 to 100 req/sec; optional, from semanticscholar.org/product/api.
BIOGRID_API_KEY optionalsecret BioGRID key for protein-protein interaction queries; optional, free registration at webservice.thebiogrid.org.
DISGENET_API_KEY optionalsecret DisGeNET key for gene-disease association data; optional, free registration at disgenet.org.
OMIM_API_KEY optionalsecret OMIM key for Mendelian disease data; optional, register at omim.org/api.
ONCOKB_API_TOKEN optionalsecret OncoKB token for precision oncology annotations; optional, register at oncokb.org/apiAccess.
NVIDIA_API_KEY optionalsecret NVIDIA NIM key for AlphaFold2 structure prediction, molecular docking, and genomics tools; optional, from build.nvidia.com.
USPTO_API_KEY optionalsecret USPTO key for patent data access; optional, register at developer.uspto.gov.
UMLS_API_KEY optionalsecret UMLS key for medical terminology and concept mapping; optional, register at uts.nlm.nih.gov/uts/.
BRENDA_PASSWORD optionalsecret BRENDA enzyme database login password; optional.
Other optional settings (2)
BRENDA_EMAIL optional BRENDA enzyme database login email; optional, register at brenda-enzymes.org.
PYTHONIOENCODING optional Set to utf-8 in the client config to avoid Python output encoding issues; optional.

The recommended setup is to ask an AI agent to read https://aiscientist.tools/setup.md and configure MCP, API keys, skills, and validation. For manual setup, add the JSON in install_config to the MCP configuration file. Agent skills can also be installed with npx skills add mims-harvard/ToolUniverse. Use --refresh to check PyPI for the latest release at launch; omit it to use uv's cache.

MCP config · JSON 
{"mcpServers":{"tooluniverse":{"command":"uvx","args":["--refresh","tooluniverse"],"env":{"PYTHONIOENCODING":"utf-8"}}}}

Works in clients that read an "mcpServers" key. For clients such as VS Code that use a different key, the configurator below converts it.

.vscode/mcp.json
{
  "servers": {
    "tooluniverse": {
      "command": "uvx",
      "args": [
        "--refresh",
        "tooluniverse"
      ],
      "env": {
        "PYTHONIOENCODING": "utf-8"
      }
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal
claude mcp add tooluniverse -e PYTHONIOENCODING=utf-8 -- uvx --refresh tooluniverse

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

After launching (uvx --refresh tooluniverse), the client's tool list should show the 4 compact-mode core tools (tool finder and execute_tool among them); ask the assistant to run any scientific tool via execute_tool (e.g., a UniProt lookup) — a returned result confirms the connection.

Troubleshooting

  1. Confirm that uv is installed and start the server with uvx or the README configuration
  2. Run `tooluniverse-doctor` to identify missing extras
  3. Use `--refresh` when a current PyPI release is needed; clear the package cache with `uv cache clean tooluniverse` if necessary
  4. Use repeated `--categories` options to limit loaded tool categories and review compact-mode behavior
  5. Check API keys, registration status, and rate limits for the external services being used

Things to try

Once connected, you can ask your AI assistant things like:

  • Use the tool finder to locate a UniProt tool, then look up the TP53 protein.
  • Search PubMed for recent review articles on CRISPR gene editing.
  • Use execute_tool to query ChEMBL for information on the drug muraglitazar.
  • Find target genes associated with Parkinson's disease using OpenTargets.

Use cases

Life-science and drug-discovery research
Literature search across PubMed, Semantic Scholar, ArXiv, BioRxiv, Europe PMC, and more
Data analysis, knowledge retrieval, and experimental design
Long-running tasks such as protein docking and molecular simulations
Composable workflows for AI scientist systems

Overview

ToolUniverse is an ecosystem for building AI scientist systems, covering life science, research, literature search, data analysis, and experimental design. Its MCP server supports compact mode, exposing a small set of core discovery and execution capabilities by default while retaining access to more than 2,500 tools.

Similar servers

BioMCP 75 · B

One binary. One grammar. Evidence from the biomedical sources you already trust.

★ 654 · Tools 11 Compare with this →
Compare with other MCP servers

Source revision 7084ad507dce Data synced 2026-10-12 Read the FMRS scoring method