Skip to content

Files

Latest commit

 

History

History
230 lines (160 loc) · 8.96 KB

README.md

File metadata and controls

230 lines (160 loc) · 8.96 KB

🤖 LangGraph Computer Use Agent (CUA)

Tip

Looking for the TypeScript version? Check out the repo here.

A Python library for creating computer use agent (CUA) systems using LangGraph. A CUA is a type of agent which has the ability to interact with a computer to preform tasks.

Short demo video:

cua-demo.mp4

Tip

This demo used the following prompt:

I want to contribute to the LangGraph.js project. Please find the GitHub repository, and inspect the read me,
along with some of the issues and open pull requests. Then, report back with a plan of action to contribute.

This library is built on top of LangGraph, a powerful framework for building agent applications, and comes with out-of-box support for streaming, short-term and long-term memory and human-in-the-loop.

Installation

pip install langgraph-cua

Quickstart

This project by default uses Scrapybara for accessing a virtual machine to run the agent. To use LangGraph CUA, you'll need both OpenAI and Scrapybara API keys.

export OPENAI_API_KEY=<your_api_key>
export SCRAPYBARA_API_KEY=<your_api_key>

Then, create the graph by importing the create_cua function from the langgraph_cua module.

from langgraph_cua import create_cua
from dotenv import load_dotenv

# Load environment variables from .env file
load_dotenv()


cua_graph = create_cua()

# Define the input messages
messages = [
    {
        "role": "system",
        "content": (
            "You're an advanced AI computer use assistant. The browser you are using "
            "is already initialized, and visiting google.com."
        ),
    },
    {
        "role": "user",
        "content": (
            "Can you find the best price for new all season tires which will fit on my 2019 Subaru Forester?"
        ),
    },
]

async def main():
    # Stream the graph execution
    stream = cua_graph.astream(
        {"messages": messages},
        stream_mode="updates"
    )

    # Process the stream updates
    async for update in stream:
        if "create_vm_instance" in update:
            print("VM instance created")
            stream_url = update.get("create_vm_instance", {}).get("stream_url")
            # Open this URL in your browser to view the CUA stream
            print(f"Stream URL: {stream_url}")

    print("Done")

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

The above example will invoke the graph, passing in a request for it to do some research into LangGraph.js from the standpoint of a new contributor. The code will log the stream URL, which you can open in your browser to view the CUA stream.

You can find more examples inside the examples directory.

How to customize

The create_cua function accepts a few configuration parameters. These are the same configuration parameters that the graph accepts, along with recursion_limit.

You can either pass these parameters when calling create_cua, or at runtime when invoking the graph by passing them to the config object.

Configuration Parameters

  • scrapybara_api_key: The API key to use for Scrapybara. If not provided, it defaults to reading the SCRAPYBARA_API_KEY environment variable.
  • timeout_hours: The number of hours to keep the virtual machine running before it times out.
  • zdr_enabled: Whether or not Zero Data Retention is enabled in the user's OpenAI account. If True, the agent will not pass the previous_response_id to the model, and will always pass it the full message history for each request. If False, the agent will pass the previous_response_id to the model, and only the latest message in the history will be passed. Default False.
  • recursion_limit: The maximum number of recursive calls the agent can make. Default is 100. This is greater than the standard default of 25 in LangGraph, because computer use agents are expected to take more iterations.
  • auth_state_id: The ID of the authentication state. If defined, it will be used to authenticate with Scrapybara. Only applies if 'environment' is set to 'web'.
  • environment: The environment to use. Default is web. Options are web, ubuntu, and windows.
  • prompt: The prompt to pass to the model. This will be passed as the system message.

System Prompts

Including a system prompt with your CUA graph is recommended, and can save the agent time in its initial steps by providing context into its environment and objective. Below is the recommended system prompt from Scrapybara:

System Prompt 
You have access to an Ubuntu VM with internet connectivity. You can install Ubuntu applications using the bash tool (prefer curl over wget).  

### Handling HTML and Large Text Output  
- To read an HTML file, open it in Chromium using the address bar.  

### Interacting with Web Pages and Forms  
- Zoom out or scroll to ensure all content is visible.  
- When interacting with input fields:  
- Clear the field first using `Ctrl+A` and `Delete`.  
- Take an extra screenshot after pressing "Enter" to confirm the input was submitted correctly.  
- Move the mouse to the next field after submission.  

### Efficiency and Authentication  
- Computer function calls take time; optimize by stringing together related actions when possible.  
- You are allowed to take actions on authenticated sites on behalf of the user.  
- Assume the user has already authenticated if they request access to a site.  
- For logging into additional sites, ask the user to use Auth Contexts or the Interactive Desktop.  

### Handling Black Screens  
- If the first screenshot shows a black screen:  
- Click the center of the screen.  
- Take another screenshot.  

### Best Practices  
- If given a complex task, break it down into smaller steps and ask for details only when necessary.  
- Read web pages thoroughly by scrolling down until sufficient information is gathered.  
- Explain each action you take and why.  
- Avoid asking for confirmation on routine actions (e.g., pressing "Enter" after typing a URL). Seek clarification only for ambiguous or critical actions (e.g., deleting files or submitting sensitive information).  
- If a user's request implies the need for external information, assume they want you to search for it and provide the answer directly.  

### Date Context  
Today's date is {todays_date}

If you choose to use this prompt, ensure you're populating the {todays_date} placeholder with the current date.

Auth States

LangGraph CUA integrates with Scrapybara's auth states API to persist browser authentication sessions. This allows you to authenticate once (e.g., logging into Amazon) and reuse that session in future runs.

Using Auth States

Pass an auth_state_id when creating your CUA graph:

from langgraph_cua import create_cua

cua_graph = create_cua(auth_state_id="<your_auth_state_id>")

The graph stores this ID in the authenticated_id state field. If you change the auth_state_id in future runs, the graph will automatically reauthenticate.

Managing Auth States with Scrapybara SDK

Save an Auth State

from scrapybara import Scrapybara

client = Scrapybara(api_key="<api_key>")
instance = client.get("<instance_id>")
auth_state_id = instance.save_auth(name="example_site").auth_state_id

Modify an Auth State

client = Scrapybara(api_key="<api_key>")
instance = client.get("<instance_id>")
instance.modify_auth(auth_state_id="your_existing_auth_state_id", name="renamed_auth_state")

Note

To apply changes to an auth state in an existing run, set the authenticated_id state field to None to trigger re-authentication.

Zero Data Retention (ZDR)

LangGraph CUA supports Zero Data Retention (ZDR) via the zdr_enabled configuration parameter. When set to true, the graph will not assume it can use the previous_message_id, and all AI & tool messages will be passed to the OpenAI on each request.

Development

To get started with development, first clone the repository:

git clone https://github.com/langchain-ai/langgraph-cua.git

Create a virtual environment:

uv venv

Activate it:

source .venv/bin/activate

Then, install dependencies:

uv sync --all-groups

Next, set the required environment variables:

cp .env.example .env

Finally, you can then run the integration tests:

pytest -xvs tests/integration/test_cua.py