Cua ("koo-ah") is an open-source framework for Computer-Use Agents - enabling AI systems to autonomously operate computers through visual understanding and action execution. Used for research, evaluation, and production deployment of desktop, browser, and mobile automation agents.

Computer-Use Agents (CUAs) are AI systems that can autonomously interact with computer interfaces through visual understanding and action execution. Unlike traditional automation tools that rely on brittle selectors or APIs, CUAs use vision-language models to perceive screen content and reason about interface interactions - enabling them to adapt to UI changes and handle complex, multi-step workflows across applications.

With the Computer SDK, you can:

• automate Windows, Linux, and macOS VMs with a consistent, pyautogui-like API
• create & manage VMs locally or using Cua cloud

With the Agent SDK, you can:

• run computer-use models with a consistent schema
• benchmark on OSWorld-Verified (369 tasks), SheetBench-V2, and ScreenSpot with a single line of code using HUD - see benchmark results (Notebook)
• combine UI grounding models with any LLM using composed agents
• use new UI agent models and UI grounding models from the Model Zoo below with just a model string (e.g., ComputerAgent(model="openai/computer-use-preview"))
• use API or local inference by changing a prefix (e.g., openai/, openrouter/, ollama/, huggingface-local/, mlx/, etc.)

• Clone a starter template and run the code in <1 min
• Get started with the Cua SDKs
• Get started with the Cua CLI

Install the agent SDK:

pip install cua-agent[all]

Initialize a computer agent using a model configuration string and a computer instance:

from agent import ComputerAgent

# ComputerAgent works with any computer initialized with the Computer SDK

agent = ComputerAgent(
    model="anthropic/claude-sonnet-4-5-20250929",
    tools=[computer],
    max_trajectory_budget=5.0
)

messages = [{"role": "user", "content": "Take a screenshot and tell me what you see"}]

async for result in agent.run(messages):
    for item in result["output"]:
        if item["type"] == "message":
            print(item["content"][0]["text"])

Cua uses the OpenAI Agent response format.

{
  "output": [
    {
      "role": "user",
      "content": "go to trycua on gh"
    },
    {
      "summary": [
        {
          "text": "Searching Firefox for Trycua GitHub",
          "type": "summary_text"
        }
      ],
      "type": "reasoning"
    },
    {
      "action": {
        "text": "Trycua GitHub",
        "type": "type"
      },
      "call_id": "call_QI6OsYkXxl6Ww1KvyJc4LKKq",
      "status": "completed",
      "type": "computer_call"
    },
    {
      "type": "computer_call_output",
      "call_id": "call_QI6OsYkXxl6Ww1KvyJc4LKKq",
      "output": {
        "type": "input_image",
        "image_url": "data:image/png;base64,..."
      }
    },
    {
      "type": "message",
      "role": "assistant",
      "content": [
        {
          "text": "Success! The Trycua GitHub page has been opened.",
          "type": "output_text"
        }
      ]
    }
  ],
  "usage": {
    "prompt_tokens": 150,
    "completion_tokens": 75,
    "total_tokens": 225,
    "response_cost": 0.01
  }
}

These are the valid model configurations for ComputerAgent(model="..."):

The following table shows which capabilities are supported by each model:

Legend:

• 🖥️ Computer-Use: Full agentic loop with planning and execution
• 🎯 Grounding: UI element detection and click coordinate prediction
• 🛠️ Tools: Support for function calling beyond screen interaction
• 👁️ VLM: Vision-language understanding

Composition Examples:

See more examples on our composition docs.

# Use OpenAI's GPT-5 for planning with specialized grounding
agent = ComputerAgent(model="huggingface-local/HelloKKMe/GTA1-7B+openai/gpt-5")

# Composition via OmniParser
agent = ComputerAgent(model="omniparser+openai/gpt-4o")

# Combine state-of-the-art grounding with powerful reasoning
agent = ComputerAgent(model="huggingface-local/HelloKKMe/GTA1-7B+anthropic/claude-3-5-sonnet-20241022")

# Combine two different vision models for enhanced capabilities
agent = ComputerAgent(model="huggingface-local/ByteDance-Seed/UI-TARS-1.5-7B+openai/gpt-4o")

# Use the built-in Moondream3 grounding with any planning mode.
agent = ComputerAgent(model="moondream3+openai/gpt-4o")

Missing a model? Create a feature request or contribute!

Learn more in the Agent SDK documentation.

Install the computer SDK:

pip install cua-computer

Initialize a computer:

from computer import Computer

computer = Computer(
    os_type="linux",  # or "macos", "windows"
    provider_type="cloud",  # or "lume", "docker", "windows_sandbox"
    name="your-sandbox-name",
    api_key="your-api-key"  # only for cloud
    # or use_host_computer_server=True for host desktop
)

try:
    await computer.run()

    # Take a screenshot
    screenshot = await computer.interface.screenshot()

    # Click and type
    await computer.interface.left_click(100, 100)
    await computer.interface.type_text("Hello!")
finally:
    await computer.close()

Learn more in the Computer SDK documentation.

Install the MCP server:

pip install cua-mcp-server

Learn more in the MCP Server documentation.

Install the Computer Server:

pip install cua-computer-server
python -m computer_server

Learn more in the Computer Server documentation.

Install Lume:

curl -fsSL https://raw.githubusercontent.com/trycua/cua/main/libs/lume/scripts/install.sh | bash

Learn more in the Lume documentation.

Install Lumier:

docker pull trycua/lumier:latest

Learn more in the Lumier documentation.

Install SOM:

pip install cua-som

Learn more in the SOM documentation.

• Hack the North Competition: First benchmark-driven hackathon track with guaranteed YC interview prize. Winner achieved 68.3% on OSWorld-Tiny (Blog Post)
• Global Hackathon Launch: Ollama × Cua global online competition for creative local/hybrid agents

• v0.4 Release - Composite Agents: Mix grounding + planning models with + operator (e.g., "GTA-7B+GPT-4o") (Blog Post)
• HUD Integration: One-line benchmarking on OSWorld-Verified with live trace visualization (Blog Post)
• Human-in-the-Loop: Interactive agent mode with human/human model string
• Web-Based Computer Use: Browser-based agent execution (Blog Post)

• Windows Sandbox Support: Native Windows agent execution (Blog Post)
• Containerization Evolution: From Lume to full Docker support (Blog Post)
• Sandboxed Python Execution: Secure code execution in agent workflows

• Cua Cloud Containers: Production-ready cloud deployment with elastic scaling (Blog Post)
• Trajectory Viewer: Visual debugging tool for agent actions (Blog Post)
• Training Data Collection: Tools for creating computer-use training datasets (Blog Post)
• App-Use Framework: Mobile and desktop app automation capabilities

• Agent Framework v0.4: Unified API for 100+ model configurations
• UI-TARS Integration: Local inference support for ByteDance's desktop-optimized model
• Blog Series: "Build Your Own Operator" tutorials (Part 1 | Part 2)

• Initial Public Release: Core Agent SDK and Computer SDK
• Lume VM Manager: macOS VM management tool for local development

• Cua Blog
• Cua Docs

We welcome contributions to Cua! Please refer to our Contributing Guidelines for details.

Join our Discord community to discuss ideas, get assistance, or share your demos!

Cua is open-sourced under the MIT License - see the LICENSE file for details.

Portions of this project, specifically components adapted from Kasm Technologies Inc., are also licensed under the MIT License. See libs/kasm/LICENSE for details.

Microsoft's OmniParser, which is used in this project, is licensed under the Creative Commons Attribution 4.0 International License (CC-BY-4.0). See the OmniParser LICENSE for details.

Some optional extras for this project depend on third-party packages that are licensed under terms different from the MIT License.

• The optional "omni" extra (installed via pip install "cua-agent[omni]") installs the cua-som module, which includes ultralytics and is licensed under the AGPL-3.0.

When you choose to install and use such optional extras, your use, modification, and distribution of those third-party components are governed by their respective licenses (e.g., AGPL-3.0 for ultralytics).

Apple, macOS, and Apple Silicon are trademarks of Apple Inc.
Ubuntu and Canonical are registered trademarks of Canonical Ltd.
Microsoft is a registered trademark of Microsoft Corporation.

This project is not affiliated with, endorsed by, or sponsored by Apple Inc., Canonical Ltd., Microsoft Corporation, or Kasm Technologies.

Thank you to all our supporters!

Thank you to all our GitHub Sponsors!