Scale computer-use 2.0 with open-source drivers, cross-OS sandboxes, and benchmarks for training, evaluation, and data generation.
Rust agentai-agentapplecomputer-usecomputer-use-agentcontainerizationcuadesktop-automationhacktoberfestlumemacosmanusoperatorswiftvirtualizationvirtualization-frameworkwindowswindows-sandbox
[project]
name = "cua-bench"
version = "0.3.1"
description = "Run computer-use benchmark datasets on sandboxes"
readme = "README.md"
license = "MIT"
authors = [
{ name = "TryCua", email = "[email protected]" }
]
keywords = [
"computer-use",
"reinforcement-learning",
"benchmarks",
"gui-automation",
"agents",
"vlm",
]
classifiers = [
"Development Status :: 4 - Beta",
"Intended Audience :: Developers",
"Intended Audience :: Science/Research",
"License :: OSI Approved :: MIT License",
"Operating System :: OS Independent",
"Programming Language :: Python :: 3",
"Programming Language :: Python :: 3.12",
"Programming Language :: Python :: 3.13",
"Topic :: Scientific/Engineering :: Artificial Intelligence",
]
requires-python = ">=3.12,<3.14"
# Core dependencies - minimal set for basic functionality
dependencies = [
# Async & networking (aiohttp: the legacy computer-server client)
"aiohttp>=3.9.0",
# CLI & display
"rich>=14.2.0",
"tqdm>=4.66.0",
# Agent prompt templates
"jinja2>=3.1.0",
# `cb dataset build` (HTML snapshot parsing)
"html5lib>=1.1",
"beautifulsoup4>=4.14.2",
# Traces (HF datasets), screenshots, .env files
"pillow>=11.0.0",
"python-dotenv>=1.0.0",
"datasets>=3.0.0",
# CUA SDK. No direct cua-core pin: cua-bench's telemetry uses it when
# present, and harnesses pinned to cua-core 0.1 (cua-computer 0.5 /
# cua-agent 0.7, e.g. Agents' Last Exam) must co-install. cua-agent stays
# unpinned for the same reason; the bundled agents need cua-agent >= 0.8.
"cua-sandbox>=0.8.0,<0.10.0",
"cua-agent",
]
[project.optional-dependencies]
# Full installation with all features
all = [
"cua-bench[browser,agents,otel,spice]",
]
# Packaging scripts (e.g. package_kicad_submission.py)
packaging = [
"boto3>=1.26.0",
]
# Deprecated and empty: the simulated (Playwright) provider was removed in
# 0.3. Kept for one release so `cua-bench[browser]` still installs.
browser = []
# Agent support
agents = [
"cua-agent>=0.8.0",
"anthropic>=0.26.0",
]
# OpenTelemetry tracing and metrics export
otel = [
"opentelemetry-api>=1.20.0",
"opentelemetry-sdk>=1.20.0",
"opentelemetry-exporter-otlp-proto-http>=1.20.0",
]
# SPICE circuit simulation (for KiCad netlist evaluation)
spice = [
"PySpice>=1.5",
]
# Deprecated and empty: the env worker server moved to the cua-bench-rl
# package in 0.3 (pip install cua-bench-rl).
server = []
# Windows-specific
windows = [
"pypiwin32>=223",
"icoextract>=0.2.0",
]
# OSWorld adapter (tasks/osworld): what the pinned upstream getters and
# metrics import at module level or use for the common task domains. The
# heavy optional ones (easyocr/torch, librosa, acoustid, borb, pydrive) are
# shimmed; the few metrics that need them fail with a clear error.
osworld = [
"requests>=2.31",
"requests-toolbelt>=1.0",
"lxml>=5.0",
"cssselect>=1.2",
"openpyxl>=3.1",
"python-docx>=1.1",
"python-pptx>=1.0",
"pypdf>=4.0",
"PyPDF2>=3.0",
"pdfplumber>=0.11",
"pymupdf>=1.24",
"rapidfuzz>=3.0",
"xmltodict>=0.13",
"beautifulsoup4>=4.12",
"odfpy>=1.4",
"pandas>=2.0",
"numpy>=1.26",
"scipy>=1.11",
"pytz>=2024.1",
"pyyaml>=6.0",
"tldextract>=5.0",
"ImageHash>=4.3",
"scikit-image>=0.22",
"opencv-python-headless>=4.8",
"mutagen>=1.47",
"formulas>=1.2",
"fastdtw>=0.3",
"gymnasium>=0.28",
"python-dotenv>=1.0",
"chardet>=5.0",
"playwright>=1.40",
]
# Development
dev = [
"pytest>=8.0.0",
"pytest-asyncio>=0.23.0",
"httpx>=0.27.0",
"ruff>=0.1.0",
]
# Deprecated and empty: RL workers, the dataloader and trainers moved to the
# cua-bench-rl package in 0.3 (pip install "cua-bench-rl[torch]").
rl = []
[project.urls]
Homepage = "https://github.com/trycua/cua"
Documentation = "https://docs.trycua.com"
Repository = "https://github.com/trycua/cua"
Issues = "https://github.com/trycua/cua/issues"
[project.scripts]
cb = "cua_bench.cli.main:main"
cua-bench = "cua_bench.cli.main:main"
[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"
[tool.hatch.build]
include = [
"cua_bench/**",
"README.md",
"LICENSE",
]
exclude = [
"cua_bench/**/__pycache__",
"cua_bench/**/tests",
"**/.DS_Store",
]
[tool.hatch.build.targets.wheel]
packages = ["cua_bench"]
[tool.hatch.build.targets.sdist]
include = [
"cua_bench/**",
"README.md",
"LICENSE",
"pyproject.toml",
]
# Ruff configuration
[tool.ruff]
line-length = 100
target-version = "py312"
[tool.ruff.lint]
select = ["E", "F", "W"]
ignore = ["E501"]
# Pytest configuration
[tool.pytest.ini_options]
asyncio_mode = "auto"
testpaths = ["cua_bench/tests"]
[tool.uv.sources]
cua-agent = { path = "../python/agent", editable = true }
cua-sandbox = { path = "../python/cua-sandbox", editable = true }
[tool.uv]
constraint-dependencies = [
"cryptography>=48.0.1",
"pytest>=9.0.3",
"starlette>=1.3.1",
]
[dependency-groups]
dev = [
"pytest>=8.4.2",
"pytest-asyncio>=1.3.0",
]