WALT: Web Agents that Learn Tools

Web Agents that Learn Tools - Automatic tool discovery from websites for LLM agents

WALT enables LLM agents to automatically discover and learn reusable tools from any website. Point WALT at a website, and it will explore, understand, and generate ready-to-use tool definitions.

🚀 Quick Start

Installation

# Install uv (faster than pip)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install WALT (ideally inside a venv)
uv venv && source .venv/bin/activate
uv pip install sfr-walt
playwright install chromium

# Set up configuration
walt init  # Creates .env file for API keys

Basic Usage

# Run agent with tools
walt agent "find and return the URL of the cheapest blue kayak" \
  --tools walt-tools/classifieds/ \
  --start-url http://localhost:9980

# Discover new tools from any website
walt discover --url https://example.com

# Or generate a specific tool (faster!)
walt generate --url https://zillow.com --goal "Search for homes with filters"

# List available tools
walt list walt-tools/shopping/

# Start an MCP server
walt serve walt-tools/classifieds/ --port 8000

# Record a demonstration
walt record https://example.com --name my_tool

🐍 Python SDK

WALT can be used programmatically for tool discovery and agent execution:

# Tool discovery
from walt.tools.discovery import propose, generate
import asyncio

async def discover_tools():
    class Args:
        base_url = "https://example.com"
        output_dir = "my-tools"
        llm = "gpt-5-mini"
        planner_llm = "gpt-5-mini"
        auth_file = None  # Optional: path to Playwright storage_state.json
        max_processes = 16
        
    args = Args()
    
    # Phase 1: Discover candidates
    tools = await propose.discover_candidates(args)
    
    # Phase 2: Generate tools
    await generate.generate_tools(args, tools)

asyncio.run(discover_tools())

# Agent with tools
from walt.browser_use.custom.agent_zoo import VWA_Agent
from walt.browser_use.custom.browser import VWABrowser, BrowserConfig
from walt.browser_use import Controller
from walt.tools.discovery.register import register_tools_from_directory
from langchain_openai import ChatOpenAI

async def run_agent():
    # Setup browser and controller
    browser = VWABrowser(BrowserConfig(headless=False))
    controller = Controller()
    
    # Load tools
    register_tools_from_directory(
        controller=controller,
        tool_dir="walt-tools/classifieds/",
        llm=ChatOpenAI(model="gpt-5-mini")
    )
    
    # Create and run agent
    agent = VWA_Agent(
        task="Find the cheapest blue kayak",
        llm=ChatOpenAI(model="gpt-5-mini"),
        browser=browser,
        controller=controller,
        max_actions_per_step=30
    )
    
    await agent.run()
    await browser.close()

asyncio.run(run_agent())

📖 CLI Commands

`walt agent <task>`

Run an agent to complete a task, optionally using tools.

walt agent "find cheap apartments" --tools walt-tools/classifieds/ --start-url https://www.zillow.com
walt agent "book a flight to NYC" --llm gemini-2.5-flash --max-steps 100 --start-url https://www.google.com/flights
walt agent "search for blue kayaks" --save-gif kayak_search.gif  # Record as GIF

Key options: --tools, --llm, --headless, --max-steps, --start-url, --save-gif

Recording: Use --save-gif <path> to save the agent's browser interactions as an animated GIF with step-by-step actions overlaid.

`walt discover --url <url>`

Discover and generate tools by exploring a website.

walt discover --url https://example.com
walt discover --url http://localhost:9980 --output walt-tools/mysite
walt discover --url https://example.com --auth-file .auth/state.json
walt discover --url https://example.com --llm gpt-4o --max-processes 8

Key options: --url, --output, --llm, --auth-file, --max-processes, --force-regenerate

Note: To reproduce results on research benchmarks, see BENCHMARKS.md.

`walt generate --url <url> --goal <goal>`

Generate a specific tool without exploration (when you know what you want).

walt generate --url https://airbnb.com --goal "Search for homes available in a location for provided dates and guest details"
walt generate --url https://zillow.com --goal "View property details" -o walt-tools/zillow/
walt generate --url https://example.com --goal "Book appointment" --auth-file .auth/state.json

Key options: --url, --goal, --output, --llm, --auth-file

Use case: When you already know what tool you need and don't want to wait for exploratory discovery.

`walt record <url>`

Record a human demonstration and convert it to a tool.

walt record https://example.com --name search_products

`walt serve <tool_dir>`

Start an MCP server with your tools.

walt serve walt-tools/shopping/ --port 8000

`walt list [tool_dir]`

List discovered tools.

walt list                           # All tools
walt list walt-tools/classifieds/   # Specific directory
walt list --detailed                # Detailed table view

The examples/ directory contains detailed examples of how to use WALT, including:

01_simple_discovery.py - Simple tool discovery
02_agent_with_tools.py - Using an agent with discovered tools
03_advanced_tool_use.py - Advanced tool usage patterns

📦 Tool Format

WALT tools are JSON files with a simple structure:

{
  "name": "search_products",
  "description": "Search for products on the site",
  "inputs": {
    "query": {
      "type": "string",
      "description": "Search query",
      "required": true
    }
  },
  "steps": [
    {
      "type": "navigation",
      "url": "https://example.com"
    },
    {
      "type": "input",
      "cssSelector": "#search-box",
      "text": "{query}"
    },
    {
      "type": "click",
      "cssSelector": "#search-button"
    },
    {
      "type": "extract_page_content",
      "goal": "Extract search results"
    }
  ]
}

Step types:

Deterministic: navigation, click, input, select_change, key_press, scroll
Agentic: extract_page_content, wait_for_page_load

See walt-tools/ for 50 pre-discovered examples.

🛠️ Development

Install from Source

git clone https://github.com/salesforceairesearch/walt.git
cd walt
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"
playwright install chromium

Project Structure

walt/
├── src/walt/
│   ├── browser_use/         # Browser automation
│   ├── tools/               # Tool system (discovery, execution, demonstration)
│   ├── benchmarks/          # WebArena/VisualWebArena evaluation
│   ├── cli.py               # CLI entry point
│   └── config.py            # Configuration system
├── experiment_configs/
│   └── ...                  # Experiment & benchmark configs
├── walt-tools/              # Pre-discovered tools
└── examples/                # Example scripts

Configuration

Use experiment configs to define reproducible evaluation runs:

# experiment_configs/my_experiment.yaml
name: "My Experiment"
llm:
  agent_model: gpt-5
agent:
  max_steps: 100
output:
  dir: outputs/my-experiment

Run it: python src/walt/benchmarks/vwa/aeval.py --config experiment_configs/my_experiment.yaml

Reproducing Paper Results

Interested in reproducing results from our paper? See BENCHMARKS.md for:

WebArena and VisualWebArena setup
Running evaluations with experiment configs
Tool discovery for benchmarks
Detailed configuration options

🤝 Citation

If you use WALT in your research, please cite:

@article{walt2025,
  title={WALT: Web Agents that Learn Tools},
  author={Viraj Prabhu, Yutong Dai, Matthew Fernandez, Jing Gu, Krithika Ramakrishnan, Yanqi Luo, Silvio Savarese, Caiming Xiong, Junnan Li, Zeyuan Chen, Ran Xu},
  journal={arXiv preprint arXiv:2510.01524},
  year={2025}
}

📄 License

MIT - See LICENSE

🙏 Acknowledgments

We are grateful to the browser-use team for the following projects upon which WALT is built:

browser-use
workflow-use

We are also grateful to the WebArena and VisualWebArena teams for the benchmark datasets.

Name		Name	Last commit message	Last commit date
Latest commit History 5 Commits
experiment_configs		experiment_configs
src/walt		src/walt
walt-tools		walt-tools
.gitignore		.gitignore
BENCHMARKS.md		BENCHMARKS.md
CODEOWNERS		CODEOWNERS
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE.txt		LICENSE.txt
README.md		README.md
SECURITY.md		SECURITY.md
pyproject.toml		pyproject.toml
walt-overview.png		walt-overview.png

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Repository files navigation

WALT: Web Agents that Learn Tools

🚀 Quick Start

Installation

Basic Usage

🐍 Python SDK

📖 CLI Commands

`walt agent <task>`

`walt discover --url <url>`

`walt generate --url <url> --goal <goal>`

`walt record <url>`

`walt serve <tool_dir>`

`walt list [tool_dir]`

📦 Tool Format

🛠️ Development

Install from Source

Project Structure

Configuration

Reproducing Paper Results

🤝 Citation

📄 License

🙏 Acknowledgments

About

Uh oh!

Releases

Packages

Contributors 2

Languages

License

SalesforceAIResearch/WALT

Folders and files

Latest commit

History

Repository files navigation

WALT: Web Agents that Learn Tools

🚀 Quick Start

Installation

Basic Usage

🐍 Python SDK

📖 CLI Commands

walt agent <task>

walt discover --url <url>

walt generate --url <url> --goal <goal>

walt record <url>

walt serve <tool_dir>

walt list [tool_dir]

📦 Tool Format

🛠️ Development

Install from Source

Project Structure

Configuration

Reproducing Paper Results

🤝 Citation

📄 License

🙏 Acknowledgments

About

Resources

License

Code of conduct

Contributing

Security policy

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Contributors 2

Languages

`walt agent <task>`

`walt discover --url <url>`

`walt generate --url <url> --goal <goal>`

`walt record <url>`

`walt serve <tool_dir>`

`walt list [tool_dir]`

Packages