Discord
0Computer Use Agent Platform - AComputerComputermacOSWindowsLinuxBrowserPhoneforAIAIAgentMCPAutomationEvalsRL
Build computer use agents that see screens, click buttons, type and run code
— with Cua
Computer Use Stack
One unified interface for computer use models. Execute powerful workflows with structured outputs, multi-turn conversations, and custom tools.
# Learn more: cua.ai/docs
import os
from agent import ComputerAgent
from computer import Computer
# Set your API key from cua.ai/dashboard
os.environ["CUA_API_KEY"] = "sk_..."
# Initialize cloud computer (create from dashboard or CLI)
computer = Computer(
os_type="linux",
provider_type="cloud",
name="your-sandbox-name"
)
# Initialize agent with model and tools
agent = ComputerAgent(
model="cua/anthropic/claude-sonnet-4.5",
tools=[computer]
)
# Run agent with task
messages = [{"role": "user", "content": "Take a screenshot"}]
async for result in agent.run(messages):
print(result)Ship and Scale Fast
Cloud-powered sandboxes for your computer use agents. Simple API integration, unlimited scale.
One-line deploy
Create sandboxes, run tasks, and manage everything from your terminal
Scale Effortlessly
Run as many sandboxes as you need. No infrastructure management, no resource limits
Cross-platform
Choose between Linux, Windows, and macOS sandboxes based on your automation needs
Cloud VLM Inference
Access 100+ vision-language models from top providers with one API key
Pay-as-you-go
Only pay for what you use. Credit-based billing with no upfront costs or monthly minimums
Smart auto-routing
Automatically route requests to the best model for your task, balancing performance and cost
Open source
Agent SDK, Computer SDK, macOS virtualization, and Docker images. All open source on GitHub
Data, Train and Evaluate
Generate large-scale UI datasets, capture agent trajectories, and run standardized benchmarks. Everything you need to train and evaluate computer-use agents.
UI dataset generation
Generate diverse UI screenshots with ground-truth bounding boxes, labels, and metadata for ML training
Trajectory recording
Capture multi-step agent interactions with full state tracking. Replay, intervene, and create training data
Benchmark evals
Evaluate agents on safety, click accuracy, and multi-step tasks. From basic interactions to complex workflows
HuggingFace export
Export datasets as Arrow/Parquet for ML training. Push directly to HuggingFace Hub with one command
Multi-OS environments
Generate datasets across macOS, Windows, and Linux with authentic OS chrome, icons, and layouts
Reproducible tasks
Seeded randomization and deterministic environments. Replay from any checkpoint for debugging
Trajectory viewer
Visual replay of agent runs with step-by-step navigation. Freeze and intervene at any point
Pricing that scales with you
Pay only for what you use - credits for cloud compute and VLM inference
I am looking to...
Free
Try before you commit - perfect for testing and evaluation
- Linux Small sandbox only
- Claude Sonnet 4.5 Haiku only
- Community Discord support
- MIT License open source
Pro
Flexible credits for cloud compute and Computer-Use VLMs, with managed cloud environments and inference.
- Cloud Linux and Windows Environments
- CUA LLM Inference Provider
- Cuazar Playground UI
- Priority support via Slack
100 credits per dollar
Enterprise
Scalable cloud containers tailored for large teams and organizations.
- Everything in Pro
- 24/7 support
- HIPAA, SOC Type 1/2 Reports
What do credits buy?
Credits are our unified currency for both compute time and AI inference
How many credits do you need per month?
Sandbox Compute
Per hour for cloud environments
Billed per minute. Sessions include full desktop access with persistent storage.
VLM Inference
Per million tokens processed
Input and output tokens combined. Actual costs vary by model and usage.
Frequently Asked Questions
Everything you need to know about CUA and computer-use agents
1 minute to your first agent
No DevOps. No infrastructure. Just ship.
Agent Runs to Date
No credit card required!