Oven

The AI Browser Automation Framework
Read the Docs

Stagehand is the easiest way to build browser automations with AI-powered interactions.

• act — Instruct the AI to perform actions (e.g. click a button or scroll).

await stagehand.page.act("click on the 'Quickstart' button")

• extract — Extract and validate data from a page using a JSON schema (generated either manually or via a Pydantic model).

await stagehand.page.extract("the summary of the first paragraph")

• observe — Get natural language interpretations to, for example, identify selectors or elements from the DOM.

await stagehand.page.observe("find the search bar")

• agent — Execute autonomous multi-step tasks with provider-specific agents (OpenAI, Anthropic, etc.).

await stagehand.agent.execute("book a reservation for 2 people for a trip to the Maldives")

Installation

Install the Python package via pip:

pip install stagehand-py

Requirements

• Python 3.9+
• httpx (for async client)
• requests (for sync client)
• asyncio (for async client)
• pydantic
• python-dotenv (optional, for .env support)
• playwright
• rich (for examples/ terminal support)

You can simply run:

pip install -r requirements.txt

requirements.txt

httpx>=0.24.0
asyncio>=3.4.3 
python-dotenv>=1.0.0
pydantic>=1.10.0
playwright>=1.42.1
requests>=2.31.0
rich
browserbase

Environment Variables

Before running your script, set the following environment variables:

export BROWSERBASE_API_KEY="your-api-key"
export BROWSERBASE_PROJECT_ID="your-project-id"
export MODEL_API_KEY="your-openai-api-key"  # or your preferred model's API key
export STAGEHAND_API_URL="url-of-stagehand-server"

You can also make a copy of .env.example and add these to your .env file.

Quickstart

Stagehand supports both synchronous and asynchronous usage. Here are examples for both approaches:

Sync Client

import os
from stagehand.sync import Stagehand
from stagehand import StagehandConfig
from dotenv import load_dotenv

load_dotenv()

def main():
    # Configure Stagehand
    config = StagehandConfig(
        env="BROWSERBASE",
        api_key=os.getenv("BROWSERBASE_API_KEY"),
        project_id=os.getenv("BROWSERBASE_PROJECT_ID"),
        model_name="gpt-4o",
        model_client_options={"apiKey": os.getenv("MODEL_API_KEY")}
    )

    # Initialize Stagehand
    stagehand = Stagehand(config=config, server_url=os.getenv("STAGEHAND_API_URL"))
    stagehand.init()
    print(f"Session created: {stagehand.session_id}")

    # Navigate to a page
    stagehand.page.goto("https://google.com/")

    # Use Stagehand AI primitives
    stagehand.page.act("search for openai")

    # Combine with Playwright
    stagehand.page.keyboard.press("Enter")

    # Observe elements on the page
    observed = stagehand.page.observe("find the news button")
    if observed:
        stagehand.page.act(observed[0])  # Act on the first observed element

    # Extract data from the page
    data = stagehand.page.extract("extract the first result from the search")
    print(f"Extracted data: {data}")

    # Close the session
    stagehand.close()

if __name__ == "__main__":
    main()

Async Client

import os
import asyncio
from stagehand import Stagehand, StagehandConfig
from dotenv import load_dotenv

load_dotenv()

async def main():
    # Configure Stagehand
    config = StagehandConfig(
        env="BROWSERBASE",
        api_key=os.getenv("BROWSERBASE_API_KEY"),
        project_id=os.getenv("BROWSERBASE_PROJECT_ID"),
        model_name="gpt-4o",
        model_client_options={"apiKey": os.getenv("MODEL_API_KEY")}
    )

    # Initialize Stagehand
    stagehand = Stagehand(config=config, server_url=os.getenv("STAGEHAND_API_URL"))
    await stagehand.init()
    print(f"Session created: {stagehand.session_id}")
    
    # Get page reference
    page = stagehand.page

    # Navigate to a page
    await page.goto("https://google.com/")

    # Use Stagehand AI primitives
    await page.act("search for openai")

    # Combine with Playwright
    await page.keyboard.press("Enter")

    # Observe elements on the page
    observed = await page.observe("find the news button")
    if observed:
        await page.act(observed[0])  # Act on the first observed element

    # Extract data from the page
    data = await page.extract("extract the first result from the search")
    print(f"Extracted data: {data}")

    # Close the session
    await stagehand.close()

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

Agent Example

import os
from stagehand.sync import Stagehand
from stagehand import StagehandConfig
from stagehand.schemas import AgentConfig, AgentExecuteOptions, AgentProvider
from dotenv import load_dotenv

load_dotenv()

def main():
    # Configure Stagehand
    config = StagehandConfig(
        env="BROWSERBASE",
        api_key=os.getenv("BROWSERBASE_API_KEY"),
        project_id=os.getenv("BROWSERBASE_PROJECT_ID"),
        model_name="gpt-4o",
        model_client_options={"apiKey": os.getenv("MODEL_API_KEY")}
    )

    # Initialize Stagehand
    stagehand = Stagehand(config=config, server_url=os.getenv("STAGEHAND_API_URL"))
    stagehand.init()
    print(f"Session created: {stagehand.session_id}")
    
    # Navigate to Google
    stagehand.page.goto("https://google.com/")
    
    # Configure the agent
    agent_config = AgentConfig(
        provider=AgentProvider.OPENAI,
        model="computer-use-preview",
        instructions="You are a helpful web navigation assistant. You are currently on google.com."
        options={"apiKey": os.getenv("MODEL_API_KEY")}
    )
    
    # Define execution options
    execute_options = AgentExecuteOptions(
        instruction="Search for 'latest AI news' and extract the titles of the first 3 results",
        max_steps=10,
        auto_screenshot=True
    )
    
    # Execute the agent task
    agent_result = stagehand.agent.execute(agent_config, execute_options)
    
    print(f"Agent execution result: {agent_result}")
    
    # Close the session
    stagehand.close()

if __name__ == "__main__":
    main()

Pydantic Schemas

• ActOptions

  The ActOptions model takes an action field that tells the AI what to do on the page, plus optional fields such as useVision and variables:

  from stagehand.schemas import ActOptions
  
  # Example:
  await page.act(ActOptions(action="click on the 'Quickstart' button"))
  

• ObserveOptions

  The ObserveOptions model lets you find elements on the page using natural language. The onlyVisible option helps limit the results:

  from stagehand.schemas import ObserveOptions
  
  # Example:
  await page.observe(ObserveOptions(instruction="find the button labeled 'News'", onlyVisible=True))
  

• ExtractOptions

  The ExtractOptions model extracts structured data from the page. Pass your instructions and a schema defining your expected data format. Note: If you are using a Pydantic model for the schema, call its .model_json_schema() method to ensure JSON serializability.

  from stagehand.schemas import ExtractOptions
  from pydantic import BaseModel
  
  class DescriptionSchema(BaseModel):
      description: str
  
  # Example:
  data = await page.extract(
      ExtractOptions(
          instruction="extract the description of the page",
          schemaDefinition=DescriptionSchema.model_json_schema()
      )
  )
  description = data.get("description") if isinstance(data, dict) else data.description
  

Actions caching

You can cache actions in Stagehand to avoid redundant LLM calls. This is particularly useful for actions that are expensive to run or when the underlying DOM structure is not expected to change.

Using observe to preview an action

observe lets you preview an action before taking it. If you are satisfied with the action preview, you can run it in page.act with no further LLM calls.

# Get the action preview
action_preview = await page.observe("Click the quickstart link")

# action_preview is a JSON-ified version of a Playwright action:
# {
#     "description": "The quickstart link",
#     "action": "click",
#     "selector": "/html/body/div[1]/div[1]/a",
#     "arguments": []
# }

# NO LLM INFERENCE when calling act on the preview
await page.act(action_preview[0])

Simple caching

Here's an example of implementing a simple file-based cache:

import json
from pathlib import Path
from typing import Optional, Dict, Any

# Get the cached value (None if it doesn't exist)
async def get_cache(key: str) -> Optional[Dict[str, Any]]:
    try:
        cache_path = Path("cache.json")
        if not cache_path.exists():
            return None
        with open(cache_path) as f:
            cache = json.load(f)
            return cache.get(key)
    except Exception:
        return None

# Set the cache value
async def set_cache(key: str, value: Dict[str, Any]) -> None:
    cache_path = Path("cache.json")
    cache = {}
    if cache_path.exists():
        with open(cache_path) as f:
            cache = json.load(f)
    cache[key] = value
    with open(cache_path, "w") as f:
        json.dump(cache, f)

Act with cache

Here's a function that checks the cache, gets the action, and runs it:

async def act_with_cache(page, key: str, prompt: str):
    # Check if we have a cached action
    cached_action = await get_cache(key)
    
    if cached_action:
        # Use the cached action
        action = cached_action
    else:
        # Get the observe result (the action)
        action = await page.observe(prompt)
        # Cache the action
        await set_cache(key, action[0])
    
    # Run the action (no LLM inference)
    await page.act(action[0])

You can now use act_with_cache to run an action with caching:

prompt = "Click the quickstart link"
key = prompt  # Simple cache key
await act_with_cache(page, key, prompt)

Why?

Stagehand adds determinism to otherwise unpredictable agents.

While there's no limit to what you could instruct Stagehand to do, our primitives allow you to control how much you want to leave to an AI. It works best when your code is a sequence of atomic actions. Instead of writing a single script for a single website, Stagehand allows you to write durable, self-healing, and repeatable web automation workflows that actually work.

[!NOTE] Stagehand is currently available as an early release, and we're actively seeking feedback from the community. Please join our Slack community to stay updated on the latest developments and provide feedback.

Configuration

Stagehand can be configured via environment variables or through a StagehandConfig object. Available configuration options include:

• STAGEHAND_API_URL: URL of the Stagehand API server.
• browserbase_api_key: Your Browserbase API key (BROWSERBASE_API_KEY).
• browserbase_project_id: Your Browserbase project ID (BROWSERBASE_PROJECT_ID).
• model_api_key: Your model API key (e.g. OpenAI, Anthropic, etc.) (MODEL_API_KEY).
• verbose: Verbosity level (default: 1).
  • Level 0: Error logs
  • Level 1: Basic info logs (minimal, maps to INFO level)
  • Level 2: Medium logs including warnings (maps to WARNING level)
  • Level 3: Detailed debug information (maps to DEBUG level)
• model_name: Optional model name for the AI (e.g. "gpt-4o").
• dom_settle_timeout_ms: Additional time (in ms) to have the DOM settle.
• debug_dom: Enable debug mode for DOM operations.
• stream_response: Whether to stream responses from the server (default: True).
• timeout_settings: Custom timeout settings for HTTP requests.

Example using a unified configuration:

from stagehand import StagehandConfig
import os

config = StagehandConfig(
    env="BROWSERBASE" if os.getenv("BROWSERBASE_API_KEY") and os.getenv("BROWSERBASE_PROJECT_ID") else "LOCAL",
    api_key=os.getenv("BROWSERBASE_API_KEY"),
    project_id=os.getenv("BROWSERBASE_PROJECT_ID"),
    debug_dom=True,
    headless=False,
    dom_settle_timeout_ms=3000,
    model_name="gpt-4o-mini",
    model_client_options={"apiKey": os.getenv("MODEL_API_KEY")},
    verbose=3  # Set verbosity level: 1=minimal, 2=medium, 3=detailed logs
)

License

MIT License (c) 2025 Browserbase, Inc.