LiteWebAgent

Disclaimer: Please note that LiteWebAgent is not affiliated with any for-profit company. This is a collaborative project from PathOnAI.org, an open-source AI research community, where Danqing Zhang (danqing.zhang.personal@gmail.com) is the main contributor and lead author of the NAACL paper. If anyone claims LiteWebAgent is affiliated with any for-profit company, please contact Danqing Zhang (danqing.zhang.personal@gmail.com) for verification.

Follow-Up Work: VisualTreeSearch

VisualTreeSearch is a production-ready system for visualizing and understanding web agent test-time scaling. It builds upon LiteWebAgent to provide an intuitive framework for researchers and users to understand tree search execution in web agents.

Resources

• 🌐 Project Page: PathOnAI.org/VisualTreeSearch
• 🎮 Live Demo: visual-tree-search.pathonai.org
• 📄 Paper: Read on arXiv
• 🎥 Video Demo: Watch on YouTube

Join the community!

📰 News

• [2025-03-06] The 'LiteWebAgent: The Open-Source Suite for VLM-Based Web-Agent Applications' paper is now available on arXiv: https://arxiv.org/abs/2503.02950
• [2025-02-28] "LiteWebAgent: The Open-Source Suite for VLM-Based Web-Agent Applications" was accepted by 2025 Annual Conference of the North American Chapter of the Association for Computational Linguistics -- System Demonstration Track (NAACL 2025).
• [2025-01-09] We reused LiteWebAgent infrastructure to build LiteWebAgent Tree Search, which is set for release on 02/15.
• [2025-01-01] We built a standalone evaluation suite for X-WebArena evaluation and integrated it with LiteWebAgent.
• [2024-12-12] Released an async version of LiteWebAgent for improved compatibility with FastAPI AI backends.
• [2024-12-11] Deployed LiteWebAgent’s frontend and backend on Vercel.
• [2024-12-03] zzfoo integrated AWM (Agent Workflow Memory) into the LiteWebAgent framework.
• [2024-11-25] We set up a Chrome extension prototype using LiteWebAgent as an AI backend server to control the Chrome browser via Chrome DevTools Protocol.
• [2024-11-01] We refactored LiteWebAgent's tree search into a new repository called LLMWebAgentTreeSearch.
• [2024-10-01] Completed a major refactoring of LiteWebAgent to make it flexible for importing the package, enabling the addition of web browsing capabilities to any AI agent.
• [2024-09-20] We reimplemented the paper Tree Search for Language Model Agents in the LiteWebAgent framework. Now, the search agent is capable of exploring different trajectories for accomplishing web browsing tasks and returning the most promising one. This is useful for finding the optimal path to complete complex web browsing tasks in an offline manner.
• [2024-08-22] The initial version of LiteWebAgent was released, providing a robust framework for using natural language to control a web agent.

1. QuickStart

From PyPI: https://pypi.org/project/litewebagent/

pip install litewebagent 

Then, a required step is to setup playwright by running

playwright install chromium

Test playwright & chromium installation by running this script

python test_installation.py 

Then please create a .env file, and update your API keys:

cp .env.example .env

You are ready to go! Try FunctionCallingAgent on google.com

python examples/google_test.py

2. Development mode

(1) Installation

Set up locally

First set up virtual environment, and allow your code to be able to see 'litewebagent'

python3.11 -m venv venv
. venv/bin/activate
pip3.11 install -e .

Then please create a .env file, and update your API keys:

cp .env.example .env

Test playwright & chromium installation by running this script

python3.11 test_installation.py 

(2) Try different agents

• use prompting-based web agent to finish some task and save the workflow

## easy case
python3.11 -m prompting_main --agent_type PromptAgent --starting_url https://www.google.com --goal 'search dining table' --plan 'search dining table' --log_folder log
## more complicated case
python3.11 -m prompting_main --agent_type PromptAgent --starting_url https://www.amazon.com/ --goal 'add a bag of dog food to the cart.' --plan 'add a bag of dog food to the cart.' --log_folder log

• we also provide function-calling-based web agent

## easy case
python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.google.com --goal 'search dining table' --plan 'search dining table' --log_folder log
python3.11 -m function_calling_main --agent_type HighLevelPlanningAgent --starting_url https://www.google.com --goal 'search dining table' --plan 'search dining table' --log_folder log
python3.11 -m function_calling_main --agent_type ContextAwarePlanningAgent --starting_url https://www.google.com --goal 'search dining table' --plan 'search dining table' --log_folder log
## more complicated case
python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.amazon.com/ --goal 'add a bag of dog food to the cart.' --plan 'add a bag of dog food to the cart.' --log_folder log

https://www.loom.com/share/1018bcc4e21c4a7eb517b60c2931ee3c https://www.loom.com/share/aa48256478714d098faac740239c9013 https://www.loom.com/share/89f5fa69b8cb49c8b6a60368ddcba103 https://www.loom.com/share/8c59dc1a6f264641b6a448fb6b7b4a5c

(3) test different input features

We use axtree by default. Alternatively, you can provide a comma-separated string listing the desired input feature types.

python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.airbnb.com --goal 'set destination as San Francisco, then search the results' --plan '(1) enter the "San Francisco" as destination, (2) and click search' --log_folder log
python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.airbnb.com --goal 'set destination as San Francisco, then search the results' --plan '(1) enter the "San Francisco" as destination, (2) and click search' --features interactive_elements --log_folder log
python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.airbnb.com --goal 'set destination as San Francisco, then search the results' --plan '(1) enter the "San Francisco" as destination, (2) and click search' --features axtree,interactive_elements --log_folder log

(4) auto login

First, tell Git to ignore future changes to state.json:

git update-index --skip-worktree state.json

Then run the load_state.py script and log into the websites to enable auto-login:

python3.11 load_state.py save

(5) memory

We integrated AWM (Agent Workflow Memory) into the LiteWebAgent framework. You can follow these three steps to include induced workflows as memory for the web agent, we use 'add a bag of dog food to the cart' on amazon website as an example:

Step 1: Induce workflows from mind2web datasets

python3.11 memory/mind2web_workflows_induction.py --websites amazon

Please note that you can induce workflows for multiple websites by passing a comma-separated list of website names to the --websites parameter:

python3.11 memory/mind2web_workflows_induction.py --websites amazon,aa

Step 2: Embed and store workflows in DB for retrieval

python3.11 memory/update_vector_store.py

Step 3: Run function calling agent with memory

python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.amazon.com/ --goal 'add a bag of dog food to the cart.' --workflow_memory_website amazon

(6) Use LiteWebAgent AI backend

Start the Python backend server:

python3.11 -m api.server --port 5001

3. Paper reimplementation

| Paper | Agent | |--------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------| | SoM (Set-of-Mark) Agent | PromptAgent | | Mind2Web