mims-harvard__tooluniverse/pyproject.toml

[build-system]
requires = ["setuptools", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "tooluniverse"
version = "1.5.0"
description = "A comprehensive collection of scientific tools for Agentic AI, offering integration with the ToolUniverse SDK and MCP Server to support advanced scientific workflows."
authors = [
    { name = "Shanghua Gao", email = "shanghuagao@gmail.com" }
]
dependencies = [
    "requests>=2.32.0",
    "boltz-api>=0.46.0,<0.47.0",
    "numpy>=2.2.0",
    "sympy>=1.12.0", # Symbolic mathematics for Python executor
    "graphql-core>=3.2.0",
    "fastapi>=0.116.0",
    "uvicorn>=0.36.0",
    "pydantic>=2.11.0",
    "epam.indigo>=1.34.0",
    "networkx>=3.4.0",
    # openai / google-genai are NOT here on purpose: they are optional extras
    # (`tooluniverse[openai]`, `tooluniverse[gemini]`), both covered by [all].
    "pyyaml>=6.0.0",
    "mcp[cli]>=1.29.0,<2.0.0",
    "fastmcp>=3.4.5,<4.0.0",
    "xmltodict>=1.0.0",
    "lxml>=6.0.0",
    "huggingface_hub>=0.34.0",
    "jsonpath-ng>=1.6.0",
    "rcsb-api>=1.4.0",
    # Imported at module scope by mcp_client_tool.py for the websocket
    # transport. It reached installs only as a transitive of fastmcp's
    # `fastmcp-slim[server]` extra, so a reshuffle upstream would have broken
    # the client with no local change. No upper bound on purpose: the
    # google-genai cap is what pushes downstream resolvers onto stale
    # tooluniverse releases (issue #526), and that stays inside its extra.
    "websockets>=13.0",
    # `from urllib3.util.retry import Retry` in uspto_tool.py and
    # `urllib3.disable_warnings` in euhealth/euhealth_live.py. requests already
    # requires urllib3>=1.26,<3, so this adds no new constraint; it just stops
    # a direct import from depending on another package's dependency.
    "urllib3>=1.26",
    # Imported by dose_response, survival, nca, drug_synergy, metaboanalyst,
    # timer and two remote tools. It was previously reaching installs only as a
    # transitive of that `fitz` placeholder (which required nibabel, nipype,
    # pyxnat, scipy and more), so removing it left these modules without their
    # dependency. It stays listed in the graph/visualization extras as well.
    "scipy>=1.7.0",
    "pandas>=2.2.3",
    "openpyxl>=3.1.0", # Read .xlsx bulk downloads (CellMarker, dataset tools)
    "setuptools>=70.0.0", # Fix pkg_resources deprecation warnings
    "pdfplumber>=0.11.0",
    "playwright>=1.55.0",
    "faiss-cpu==1.12.0",
    "flask>=2.0.0",
    "aiohttp",
    "beautifulsoup4>=4.12.0",
    "python-dotenv>=1.0.0",
    # MarkItDown (~32 MB: SpeechRecognition, youtube-transcript-api, pydub,
    # azure-ai-documentintelligence). The broad converter set is deliberate:
    # unified_guideline_tools.py imports markitdown at module scope and
    # convert_to_markdown takes an unconstrained URI with no format allowlist.
    #
    # These are MarkItDown's own converter extras listed one by one instead of
    # its `all` extra, and that difference is load-bearing (issue #527):
    #   * `all` pins youtube-transcript-api~=1.0.0, and every 1.0.x release
    #     declares Requires-Python <3.14, which transitively capped the Python
    #     versions tooluniverse could be installed on. The `youtube-transcription`
    #     extra leaves that dependency unpinned, so 1.2.3+ (Requires-Python
    #     <3.15) resolves and the cap is gone.
    #   * Since markitdown 0.1.6 `all` also requires
    #     azure-ai-contentunderstanding>=1.2.0b1, which has no stable release on
    #     PyPI, so default resolvers silently backtrack `markitdown[all]` to
    #     0.1.5. Naming the extras keeps installs on current MarkItDown.
    # Add `az-content-understanding` back once that dependency ships a stable
    # release; adding it today makes the whole requirement unresolvable.
    "markitdown[audio-transcription,az-doc-intel,docx,outlook,pdf,pptx,xls,xlsx,youtube-transcription]>=0.1.0",
    "psutil>=5.9.0",
    "ddgs>=9.0.0",
    "pip>=25.3",
    "jsonschema>=4.23.0",
]
requires-python = ">=3.10"
readme = "README.md"

[project.optional-dependencies]
pdf = [
    # PyMuPDF is AGPL-3.0 or commercially licensed. Keep it opt-in and out of
    # the aggregate `all` extra so default Apache-2.0 installations do not pull
    # it in implicitly.
    "pymupdf>=1.24.3",
]
ocr = [
    # Scanned-PDF text extraction for the USPTO downloader. EasyOCR pulls in
    # PyTorch, a very large addition none of the other 600+ tools need, so keep
    # it opt-in like `pdf` and out of the aggregate `all` extra.
    "easyocr>=1.7",
    "python-docx>=1.1",
]
# LLM provider SDKs. Both are imported lazily inside the client constructors in
# llm_clients.py (AzureOpenAIClient, OpenAICompatibleClient, OpenRouterClient,
# VLLMClient, GeminiClient) and in database_setup/embedder.py — never at module
# scope — so tooluniverse imports and runs fine without them, and a constructor
# reached without its extra raises a clear RuntimeError. Unlike `pdf` these ARE
# in the aggregate `all` extra, so `tooluniverse[all]` keeps today's behavior.
# They are separate extras rather than one combined group so that an
# OpenAI-only user is not subjected to google-genai's `websockets` upper bound,
# which is the constraint that can push a downstream universal resolver onto a
# much older tooluniverse instead of reporting a conflict (issue #526).
openai = [
    "openai>=1.107.0",
]
gemini = [
    "google-genai>=1.36.0",
]
client = [
    "requests>=2.32.0",
    "pydantic>=2.11.0",
]
smolagents = [
    "smolagents>=1.22.0",
    "gradio>=4.0.0",  # Required for GradioUI web interface
]
# Single-cell and spatial transcriptomics tools
singlecell = [
    "cellxgene-census>=1.10.0",  # CELLxGENE Census API
    "tiledbsoma>=1.15.3",  # Required for spatial data
]
dev = [
    "pytest>=7.0",
    "pytest-cov>=4.0",
    "pytest-timeout>=2.3.1",
    "pytest-mock>=3.14.0",
    "pytest-asyncio>=0.21.0",
    "pre-commit>=2.0",
    "pytest-xdist>=3.0",
    "pytest-html>=3.0",
    "requests-mock>=1.12.1",
    "ruff>=0.14.5",
    # tests/unit/test_gemini_client.py constructs a real google.genai Client, so
    # the provider SDKs stay part of `[dev]` now that they left the base set.
    # This keeps `pip install -e .[dev] && pytest` green, as CI runs exactly that.
    "tooluniverse[openai,gemini]",
]
docs = [
    "sphinx>=4.0",
    "furo>=2024.8.6",
    "pydata-sphinx-theme>=0.15.0",
    "sphinx-rtd-theme>=1.0",
    "shibuya>=0.10.0",
    "myst-parser>=0.18",
    "linkify-it-py>=2.0.0",
    "sphinx-autobuild>=2021.3.14",
    "sphinx-copybutton>=0.5.0",
    "sphinx-tabs>=3.5.0",
    "sphinx-design>=0.3.0",
    "sphinx-notfound-page>=0.8",
    "sphinx-autodoc-typehints>=1.12.0",
    "sphinx-reredirects>=0.1.5",
]
embedding = [
    "sentence-transformers>=5.1.0",
    "faiss-cpu==1.12.0",
    "huggingface_hub>=0.34.0",
]
ml = [
    "admet-ai>=1.2.0",
    "sentence-transformers>=5.1.0",
    "faiss-cpu==1.12.0",
    "huggingface_hub>=0.34.0",
]
graph = [
    "flask>=2.0.0",
    "matplotlib>=3.5.0",
    "plotly>=5.0.0",
    "scipy>=1.7.0",
    "pydot>=1.4.0",
    "pygraphviz>=1.7",
    "jinja2>=3.0.0",
    "werkzeug>=2.0.0",
]
visualization = [
    "py3Dmol>=2.0.0",
    "rdkit>=2023.9.1",
    "plotly>=5.18.0",
    "kaleido>=0.2.1",
    "scipy>=1.7.0",
    "matplotlib>=3.5.0",
    "networkx>=3.4.0",
]
space = [
    "huggingface_hub>=0.34.0",
    "pyyaml>=6.0.0",
    "requests>=2.32.0",
]
bioinformatics = [
    "biopython>=1.80",
    "freesasa>=2.2.0",
]
# circuit: install manually with `pip install git+https://github.com/tooluniverse/circuit.git`
all = [
    "tooluniverse[dev,docs,graph,visualization,space,embedding,ml,bioinformatics,openai,gemini]",
]
build = [
    "pyinstaller>=6.0.0",
]

[project.urls]
Homepage = "https://github.com/mims-harvard/ToolUniverse"

[project.scripts]
tu = "tooluniverse_cli_entry:main"
tooluniverse = "tooluniverse.smcp_server:run_default_stdio_server"
tooluniverse-mcp = "tooluniverse.smcp_server:run_http_server"
tooluniverse-smcp = "tooluniverse.smcp_server:run_smcp_server"
tooluniverse-smcp-server = "tooluniverse.smcp_server:run_http_server"
tooluniverse-smcp-stdio = "tooluniverse.smcp_server:run_stdio_server"
# HTTP API Server for ToolUniverse class methods
tooluniverse-http-api = "tooluniverse.http_api_server_cli:run_http_api_server"
generate-mcp-tools = "tooluniverse.generate_mcp_tools:run_generate"
# Human Expert Feedback System
tooluniverse-expert-feedback = "tooluniverse.remote.expert_feedback.human_expert_mcp_tools:main"
tooluniverse-expert-feedback-web = "tooluniverse.remote.expert_feedback.start_web_interface:main"
tu-datastore = "tooluniverse.database_setup.cli:main"
# Literature Search Web UIs (Under testing)
# tooluniverse-regular-search = "tooluniverse.web_tools.literature_search_ui.regular_search_app:main"
# tooluniverse-agent-search = "tooluniverse.web_tools.literature_search_ui.agent_search_app:main"
# tooluniverse-wide-research = "tooluniverse.web_tools.literature_search_ui.simple_app:main"
# Health Check Tool
tooluniverse-doctor = "tooluniverse.doctor:main"

[tool.setuptools]
py-modules = ["tooluniverse_cli_entry"]

[tool.setuptools.packages.find]
where = ["src"]

[tool.setuptools.package-data]
tooluniverse = ["data/*", "data/packages/*"]

[dependency-groups]
dev = [
    "tooluniverse[dev,embedding]",
]

[tool.uv.sources]
tooluniverse = { workspace = true }

[tool.ruff]
line-length = 88
exclude = [
    ".venv",
    "docs",
    "examples",
    "skills",
    "temp_docs_and_tests",
]
unsafe-fixes = true

[tool.ruff.lint]
ignore = ["E203", "E402", "E501", "F401", "F541"]