cua-agent

cua-agent is a general Computer-Use framework with liteLLM integration for running agentic workflows on macOS, Windows, and Linux sandboxes. It provides a unified interface for computer-use agents across multiple LLM providers with advanced callback system for extensibility.

Features

Safe Computer-Use/Tool-Use: Using Computer SDK for sandboxed desktops
Multi-Agent Support: Anthropic Claude, OpenAI computer-use-preview, UI-TARS, Omniparser + any LLM
Multi-API Support: Take advantage of liteLLM supporting 100+ LLMs / model APIs, including local models (huggingface-local/, ollama_chat/, mlx/)
Cross-Platform: Works on Windows, macOS, and Linux with cloud and local computer instances
Extensible Callbacks: Built-in support for image retention, cache control, PII anonymization, budget limits, and trajectory tracking

Install

pip install "cua-agent[all]"

# or install specific providers
pip install "cua-agent[openai]"        # OpenAI computer-use-preview support
pip install "cua-agent[anthropic]"     # Anthropic Claude support
pip install "cua-agent[omni]"          # Omniparser + any LLM support
pip install "cua-agent[uitars]"        # UI-TARS
pip install "cua-agent[uitars-mlx]"    # UI-TARS + MLX support
pip install "cua-agent[uitars-hf]"     # UI-TARS + Huggingface support
pip install "cua-agent[glm45v-hf]"     # GLM-4.5V + Huggingface support
pip install "cua-agent[ui]"            # Gradio UI support

Quick Start

import asyncio
import os
from agent import ComputerAgent
from computer import Computer

async def main():
    # Set up computer instance
    async with Computer(
        os_type="linux",
        provider_type="cloud",
        name=os.getenv("CUA_CONTAINER_NAME"),
        api_key=os.getenv("CUA_API_KEY")
    ) as computer:
        
        # Create agent
        agent = ComputerAgent(
            model="anthropic/claude-3-5-sonnet-20241022",
            tools=[computer],
            only_n_most_recent_images=3,
            trajectory_dir="trajectories",
            max_trajectory_budget=5.0  # $5 budget limit
        )
        
        # Run agent
        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"])

if __name__ == "__main__":
    asyncio.run(main())

Supported Models

Anthropic Claude (Computer Use API)

model="anthropic/claude-3-5-sonnet-20241022"
model="anthropic/claude-3-7-sonnet-20250219"
model="anthropic/claude-opus-4-20250514"
model="anthropic/claude-sonnet-4-20250514"

OpenAI Computer Use Preview

model="openai/computer-use-preview"

UI-TARS (Local or Huggingface Inference)

model="huggingface-local/ByteDance-Seed/UI-TARS-1.5-7B"
model="ollama_chat/0000/ui-tars-1.5-7b"

Omniparser + Any LLM

model="omniparser+ollama_chat/mistral-small3.2"
model="omniparser+vertex_ai/gemini-pro"
model="omniparser+anthropic/claude-3-5-sonnet-20241022"
model="omniparser+openai/gpt-4o"

Custom Tools

Define custom tools using decorated functions:

from computer.helpers import sandboxed

@sandboxed()
def read_file(location: str) -> str:
    """Read contents of a file
    
    Parameters
    ----------
    location : str
        Path to the file to read
        
    Returns
    -------
    str
        Contents of the file or error message
    """
    try:
        with open(location, 'r') as f:
            return f.read()
    except Exception as e:
        return f"Error reading file: {str(e)}"

def calculate(a: int, b: int) -> int:
    """Calculate the sum of two integers"""
    return a + b

# Use with agent
agent = ComputerAgent(
    model="anthropic/claude-3-5-sonnet-20241022",
    tools=[computer, read_file, calculate]
)

Callbacks System

agent provides a comprehensive callback system for extending functionality:

Built-in Callbacks

from agent.callbacks import (
    ImageRetentionCallback,
    TrajectorySaverCallback, 
    BudgetManagerCallback,
    LoggingCallback
)

agent = ComputerAgent(
    model="anthropic/claude-3-5-sonnet-20241022",
    tools=[computer],
    callbacks=[
        ImageRetentionCallback(only_n_most_recent_images=3),
        TrajectorySaverCallback(trajectory_dir="trajectories"),
        BudgetManagerCallback(max_budget=10.0, raise_error=True),
        LoggingCallback(level=logging.INFO)
    ]
)

Custom Callbacks

from agent.callbacks.base import AsyncCallbackHandler

class CustomCallback(AsyncCallbackHandler):
    async def on_llm_start(self, messages):
        """Preprocess messages before LLM call"""
        # Add custom preprocessing logic
        return messages
    
    async def on_llm_end(self, messages):
        """Postprocess messages after LLM call"""
        # Add custom postprocessing logic
        return messages
    
    async def on_usage(self, usage):
        """Track usage information"""
        print(f"Tokens used: {usage.total_tokens}")

Budget Management

Control costs with built-in budget management:

# Simple budget limit
agent = ComputerAgent(
    model="anthropic/claude-3-5-sonnet-20241022",
    max_trajectory_budget=5.0  # $5 limit
)

# Advanced budget configuration
agent = ComputerAgent(
    model="anthropic/claude-3-5-sonnet-20241022",
    max_trajectory_budget={
        "max_budget": 10.0,
        "raise_error": True,  # Raise error when exceeded
        "reset_after_each_run": False  # Persistent across runs
    }
)

Trajectory Management

Save and replay agent conversations:

agent = ComputerAgent(
    model="anthropic/claude-3-5-sonnet-20241022",
    trajectory_dir="trajectories",  # Auto-save trajectories
    tools=[computer]
)

# Trajectories are saved with:
# - Complete conversation history
# - Usage statistics and costs
# - Timestamps and metadata
# - Screenshots and computer actions

Configuration Options

ComputerAgent Parameters

model: Model identifier (required)
tools: List of computer objects and decorated functions
callbacks: List of callback handlers for extensibility
only_n_most_recent_images: Limit recent images to prevent context overflow
verbosity: Logging level (logging.INFO, logging.DEBUG, etc.)
trajectory_dir: Directory to save conversation trajectories
max_retries: Maximum API call retries (default: 3)
screenshot_delay: Delay between actions and screenshots (default: 0.5s)
use_prompt_caching: Enable prompt caching for supported models
max_trajectory_budget: Budget limit configuration

Environment Variables

# Computer instance (cloud)
export CUA_CONTAINER_NAME="your-container-name"
export CUA_API_KEY="your-cua-api-key"

# LLM API keys
export ANTHROPIC_API_KEY="your-anthropic-key"
export OPENAI_API_KEY="your-openai-key"

Advanced Usage

Streaming Responses

async for result in agent.run(messages, stream=True):
    # Process streaming chunks
    for item in result["output"]:
        if item["type"] == "message":
            print(item["content"][0]["text"], end="", flush=True)
        elif item["type"] == "computer_call":
            action = item["action"]
            print(f"\n[Action: {action['type']}]")

Interactive Chat Loop

history = []
while True:
    user_input = input("> ")
    if user_input.lower() in ['quit', 'exit']:
        break
        
    history.append({"role": "user", "content": user_input})
    
    async for result in agent.run(history):
        history += result["output"]
        
        # Display assistant responses
        for item in result["output"]:
            if item["type"] == "message":
                print(item["content"][0]["text"])

Error Handling

try:
    async for result in agent.run(messages):
        # Process results
        pass
except BudgetExceededException:
    print("Budget limit exceeded")
except Exception as e:
    print(f"Agent error: {e}")

API Reference

ComputerAgent.run()

async def run(
    self,
    messages: Messages,
    stream: bool = False,
    **kwargs
) -> AsyncGenerator[Dict[str, Any], None]:
    """
    Run the agent with the given messages.
    
    Args:
        messages: List of message dictionaries
        stream: Whether to stream the response
        **kwargs: Additional arguments
        
    Returns:
        AsyncGenerator that yields response chunks
    """

Message Format

messages = [
    {
        "role": "user",
        "content": "Take a screenshot and describe what you see"
    },
    {
        "role": "assistant", 
        "content": "I'll take a screenshot for you."
    }
]

Response Format

{
    "output": [
        {
            "type": "message",
            "role": "assistant",
            "content": [{"type": "output_text", "text": "I can see..."}]
        },
        {
            "type": "computer_call",
            "action": {"type": "screenshot"},
            "call_id": "call_123"
        },
        {
            "type": "computer_call_output",
            "call_id": "call_123",
            "output": {"image_url": "data:image/png;base64,..."}
        }
    ],
    "usage": {
        "prompt_tokens": 150,
        "completion_tokens": 75,
        "total_tokens": 225,
        "response_cost": 0.01,
    }
}

License

MIT License - see LICENSE file for details.

Metadata

Release files for cua-agent 0.4.21

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for cua-agent 0.4.21
File	Size	Uploaded
cua_agent-0.4.21.tar.gz	102.2 kB	Aug 28, 2025	Details

Built distribution (wheel)

Table of built distributions (wheels) for cua-agent 0.4.21
File	Interpreter	ABI	Platform	Reset
cua_agent-0.4.21-py3-none-any.whl 123.5 kB Aug 28, 2025	Python 3	none	any	Details

Total release size: 225.7 kB

Release files / cua_agent-0.4.21.tar.gz

Download URL	cua_agent-0.4.21.tar.gz
Size	102.2 kB
Tags	Source
SHA-256 checksum How to use checksums	`4db5f4adf1b698211e959a20b724cdaf096d3b0b245f4ee92be6c61d050a74e8`
BLAKE2b-256 checksum How to use checksums	`db5d598899616b999e889c21a1c30b5ba849ea31734fac61d7d71f0fe1e7a42c`
Upload date	Aug 28, 2025
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/6.1.0 CPython/3.11.9`

Release files / cua_agent-0.4.21-py3-none-any.whl

Download URL	cua_agent-0.4.21-py3-none-any.whl
Size	123.5 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`0e7bfced095e0d4247d873c9533637169ea1c8131d065b594cd0f94c0832438c`
BLAKE2b-256 checksum How to use checksums	`879f0dc96e3d8c5d1b271b978715811ca5e1bc6305a4e2d68fcbf4707aa56faa`
Upload date	Aug 28, 2025
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/6.1.0 CPython/3.11.9`