Migrate Browser Use to v4

Browser Use hands a task to an autonomous agent: you write Agent(task="...", llm=...), call run(), and the agent picks each action. Stagehand v4 has no equivalent object. Most of this migration hinges on these principles:

There is no Agent. Nothing in v4 takes a whole task and drives the browser for you.
v4 is a browser SDK plus three model-backed steps. act(), extract(), and observe() take natural-language instructions. Everything else is ordinary browser control.

Porting a flow means writing out the steps the agent used to infer, and deciding for each one whether it needs a model or a selector.

Browser Use is Python-first, so this guide leads with Python. Stagehand behaves the same way in TypeScript; the SDK reference carries each language’s naming.

Why there’s no Agent

Agent(task=...).run() was built for a world where a model couldn’t be trusted with the browser on its own, so the framework wrapped it in a loop, showed it the page each step, and asked it to pick one action at a time. Every step was an inference call, whether the page needed judgement or not. Many of those steps do not need a model: navigating to a URL, clicking a button with a stable selector, reading a table. v4 exposes discrete tools and leaves the control flow to you. Two approaches replace the agent:

Code mode puts the model in front of the run, not inside it. A coding assistant writes a Stagehand script once; you run that script every time after. Browserbase recommends starting here.
Tool calling keeps a model in the loop at runtime, the way Browser Use does, but drives the browser through the full Stagehand API as its tools instead of one broad task string.

Either way, act(), extract(), and observe() stay in your toolbox for the steps that genuinely need a model. You just stop handing a model the entire task.

Hello world, side by side

The smallest Browser Use program and its v4 shape:

from browser_use import Agent, ChatOpenAI
from dotenv import load_dotenv
import asyncio

load_dotenv()


async def main():
    agent = Agent(
        task="Go to news.ycombinator.com and return the title of the top story",
        llm=ChatOpenAI(model="gpt-4.1-mini"),
    )
    history = await agent.run()
    print(history.final_result())


asyncio.run(main())

import asyncio
import os

from stagehand import Stagehand, browserbase


async def main():
    browser = await browserbase.launch(api_key=os.environ["BROWSERBASE_API_KEY"])
    stagehand = await Stagehand.create(browser=browser)

    try:
        page = await browser.context.new_page("https://news.ycombinator.com")
        # A stable selector does this without a model call.
        title = await page.locator(".titleline > a").first().inner_text()
        print(title)
    finally:
        await stagehand.close()
        await browser.close()


asyncio.run(main())

The shape of the migration is already visible:

A browser factory (browserbase.launch() or local_browser.launch()) replaces the implicit browser inside Agent, and Stagehand.create() attaches the runtime to it.
The task string is gone. You write the steps.
Where Browser Use would have spent a model call reading the page, a page.locator() does it for free. Spend act() and extract() only where the page needs judgement.

Stagehand reads no environment variables of its own. Browser Use auto-loads .env and picks up OPENAI_API_KEY, BROWSER_USE_API_KEY, and friends. In v4 you pass every key explicitly: the Browserbase API key to the factory, and any model key in the model option. load_dotenv() still works to get values into os.environ; nothing reads them for you.

Code mode

Ask a coding assistant to write the Stagehand script, then run the script. The model writes the code once, instead of driving the browser on every run. You get ordinary code: reviewable, diffable, and free of per-step inference. When a site changes, re-run the assistant on the step that broke. Start with AI rules. Those rule files keep generated code on the v4 API instead of the older Stagehand and Browser Use patterns in a model’s training data. Here’s a prompt that produces a working script:

Using Stagehand v4, write a script that:
  1. Opens news.ycombinator.com
  2. Finds today's most-commented story
  3. Opens its comments and extracts the top five comment bodies

Follow the rules in my project's Stagehand rules file. Prefer page.locator()
and page.goto() for anything with a stable selector, and reserve act() and
extract() for steps that need a model.

What comes back should read like the script you’d have written yourself:

Python
TypeScript

import asyncio
import os

from pydantic import BaseModel
from stagehand import Stagehand, browserbase


class Comment(BaseModel):
    author: str
    body: str


class Comments(BaseModel):
    comments: list[Comment]


async def main() -> None:
    browser = await browserbase.launch(api_key=os.environ["BROWSERBASE_API_KEY"])
    stagehand = await Stagehand.create(browser=browser)

    try:
        page = await browser.context.new_page("https://news.ycombinator.com")

        # Deterministic where the page allows it: no inference, no variance.
        await page.locator("a.morelink").first().click()
        await page.wait_for_load_state("domcontentloaded")

        # A model call where the page needs judgement.
        await stagehand.act("Open the comments for the story with the most comments")

        result = await stagehand.extract(
            "Extract the top five comments, with each author and body",
            Comments,
        )
        print(result.data.comments)
    finally:
        await stagehand.close()
        await browser.close()


asyncio.run(main())

import { browserbase, Stagehand } from "@browserbasehq/stagehand";
import { z } from "zod/v4";

const commentSchema = z.object({
  comments: z.array(z.object({ author: z.string(), body: z.string() })),
});

const browser = await browserbase.launch({ apiKey: process.env.BROWSERBASE_API_KEY });
const stagehand = await Stagehand.create({ browser });

try {
  const page = await browser.context.newPage("https://news.ycombinator.com");

  // Deterministic where the page allows it: no inference, no variance.
  await page.locator("a.morelink").first().click();
  await page.waitForLoadState("domcontentloaded");

  // A model call where the page needs judgement.
  await stagehand.act("Open the comments for the story with the most comments");

  const { data } = await stagehand.extract(
    "Extract the top five comments, with each author and body",
    commentSchema,
  );
  console.log(data.comments);
} finally {
  await stagehand.close();
  await browser.close();
}

Generated code should use page.locator() and page.goto() wherever a selector is stable, and spend a model call only where the page needs judgement. A Browser Use agent couldn’t make that split, because every step it ran was an inference call.

Tool calling

To keep a model in the loop at runtime, the way Browser Use does, give it the whole Stagehand surface rather than one task string. Each method maps to one tool with a narrow contract, so a step names a specific browser operation instead of routing everything through one sentence of English and hoping the agent picks the right action. Browser Use also lets you register custom tools with @tools.action(...) (formerly @controller.action(...)). Those port directly: each becomes one function in the tool set below. Expose the real API:

Capability	Tools worth exposing
Navigation	`page.goto()`, `page.reload()`, `page.goBack()`, `page.goForward()`
Perception	`page.snapshot()`, `page.screenshot()`, `page.url()`, `page.title()`
Element interaction	`locator.click()`, `locator.fill()`, `locator.type()`, `locator.selectOption()`, `locator.setInputFiles()`, `locator.scrollTo()`
Element inspection	`locator.textContent()`, `locator.innerText()`, `locator.isVisible()`, `locator.isChecked()`, `locator.count()`, `locator.inputValue()`
Raw input	`page.click(x, y)`, `page.hover()`, `page.scroll()`, `page.type()`, `page.keyPress()`, `page.dragAndDrop()`
Tabs and state	`context.newPage()`, `context.pages()`, `context.setActivePage()`, `context.cookies()`
Waiting	`page.waitForSelector()`, `page.waitForLoadState()`, `page.waitForTimeout()`
Model-backed steps	`stagehand.act()`, `stagehand.extract()`, `stagehand.observe()`
Page-declared tools	`page.tools()`, see WebMCP

page.snapshot() anchors the loop the way Browser Use’s page state did. It returns formattedTree, the accessibility tree, plus an xpathMap, so the model reads real page structure and hands back a selector you can drive deterministically.

Python
TypeScript

async def goto(url: str) -> str:
    """Navigate to a URL."""
    await page.goto(url)
    return await page.url()


async def snapshot() -> str:
    """Read the accessibility tree of the current page."""
    return (await page.snapshot()).formatted_tree


async def click(selector: str) -> None:
    """Click the element matching a selector from the snapshot."""
    await page.locator(selector).click()


async def fill(selector: str, value: str) -> None:
    """Fill the input matching a selector."""
    await page.locator(selector).fill(value)


async def read_text(selector: str) -> str:
    """Read the text of the element matching a selector."""
    return await page.locator(selector).text_content()


async def act(instruction: str) -> str:
    """Perform one action in natural language when no selector is known."""
    return (await stagehand.act(instruction)).data.message


async def extract(instruction: str) -> str:
    """Read structured data off the current page."""
    return (await stagehand.extract(instruction)).data.extraction


TOOLS = [goto, snapshot, click, fill, read_text, act, extract]

import { z } from "zod/v4";

const tools = {
  goto: {
    description: "Navigate to a URL",
    parameters: z.object({ url: z.string() }),
    execute: async ({ url }: { url: string }) => {
      await page.goto(url);
      return await page.url();
    },
  },
  snapshot: {
    description: "Read the accessibility tree of the current page",
    parameters: z.object({}),
    execute: async () => (await page.snapshot()).formattedTree,
  },
  click: {
    description: "Click the element matching a selector from the snapshot",
    parameters: z.object({ selector: z.string() }),
    execute: async ({ selector }: { selector: string }) => {
      await page.locator(selector).click();
    },
  },
  fill: {
    description: "Fill the input matching a selector",
    parameters: z.object({ selector: z.string(), value: z.string() }),
    execute: async ({ selector, value }: { selector: string; value: string }) => {
      await page.locator(selector).fill(value);
    },
  },
  readText: {
    description: "Read the text of the element matching a selector",
    parameters: z.object({ selector: z.string() }),
    execute: async ({ selector }: { selector: string }) =>
      await page.locator(selector).textContent(),
  },
  act: {
    description: "Perform one action in natural language when no selector is known",
    parameters: z.object({ instruction: z.string() }),
    execute: async ({ instruction }: { instruction: string }) =>
      (await stagehand.act(instruction)).data.message,
  },
  extract: {
    description: "Read structured data off the current page",
    parameters: z.object({ instruction: z.string() }),
    execute: async ({ instruction }: { instruction: string }) =>
      (await stagehand.extract(instruction)).data.extraction,
  },
};

Escalate on observe(), never on act(). A failed act() may already have clicked, submitted, or paid before the error surfaced, so retrying it can repeat the side effect. observe() only plans, so retrying it is free. Browser Use’s max_failures retry loop had the same hazard; keep the retry on the planning step. Cost optimization applies the same idea to model escalation.

Let a coding assistant do the rest

Most of the mechanical mapping below is exactly what an assistant is good at. Point it at this page instead of retyping the rules:

Migrate this file from Browser Use to Stagehand v4.

Follow https://docs.stagehand.dev/v4/migrations/browser-use, and use its quick
reference table as the mapping. For each Agent(task=...).run() call, stop and
ask me whether to replace it with a written script (code mode) or a tool-calling
loop, and list every one you find instead of guessing.

Set up AI rules first, so the assistant stays on the v4 API instead of the Stagehand and Browser Use patterns in its training data. Then work through the sections below for anything it missed.

Recommended migration order

Get one script constructing and closing cleanly on v4, before porting any behavior. Launch a browser, open a page, close both handles.
Replace Agent(task=...).run() with code mode or tool calling. This is the real work, and everything else is mechanical.
Convert the deterministic steps to page.locator() and page.goto(), keeping act() and extract() only where the page needs judgement.
Move output_model_schema to an extract() call with a schema.
Move sensitive_data to variables on act().
Turn on server-side caching once the flow is stable.

Breaking changes

Initialization and teardown

Browser Use constructs a browser inside the agent and closes it for you. v4 separates the browser from the runtime, and you close both:

Python
TypeScript

- from browser_use import Agent, ChatOpenAI
-
- agent = Agent(task="...", llm=ChatOpenAI(model="gpt-4.1-mini"))
- await agent.run()
+ import os
+ from stagehand import Stagehand, browserbase
+
+ browser = await browserbase.launch(api_key=os.environ["BROWSERBASE_API_KEY"])
+ stagehand = await Stagehand.create(browser=browser)
+ try:
+     ...  # your steps
+ finally:
+     await stagehand.close()
+     await browser.close()

+ import { browserbase, Stagehand } from "@browserbasehq/stagehand";
+
+ const browser = await browserbase.launch({ apiKey: process.env.BROWSERBASE_API_KEY });
+ const stagehand = await Stagehand.create({ browser });
+ try {
+   // your steps
+ } finally {
+   await stagehand.close();
+   await browser.close();
+ }

Use local_browser.launch() for a browser on your machine, browserbase.launch({ apiKey }) for a hosted one, and local_browser.connect({ cdpUrl }) or browserbase.connect({ apiKey, sessionId }) to attach to one that’s already running. browserbase.launch() is what enables server-side caching and the Model Gateway. Stagehand closes only the browsers it launched, so stagehand.close() leaves the browser running and you call browser.close() yourself. See browser configuration.

The task string becomes explicit steps

There’s no direct diff here, because a task string has no single replacement. That’s the whole migration: what the agent inferred each step, you now write out. Read code mode and tool calling, pick one, and translate the intent of the task into steps.

Models

Browser Use selects a provider by the chat class you construct (ChatOpenAI, ChatAnthropic, ChatGoogle, ChatBrowserUse). v4 takes one model object, and the model name always carries a provider prefix:

Python
TypeScript

- from browser_use import Agent, ChatAnthropic
- agent = Agent(task="...", llm=ChatAnthropic(model="claude-sonnet-4-0"))
+ stagehand = await Stagehand.create(
+     browser=browser,
+     model=ModelConfig(
+         model_name="anthropic/claude-sonnet-4-6",
+         api_key=os.environ["ANTHROPIC_API_KEY"],
+     ),
+ )

+ const stagehand = await Stagehand.create({
+   browser,
+   model: {
+     modelName: "anthropic/claude-sonnet-4-6",
+     apiKey: process.env.ANTHROPIC_API_KEY,
+   },
+ });

On a Browserbase browser you can omit model entirely and the Model Gateway picks one for you, which is the closest analogue to ChatBrowserUse(). Pass the same shape to a single act(), extract(), or observe() call to override it there. Browser Use’s page_extraction_llm (a separate model for extraction) maps to passing model on the extract() call. See models.

Structured output

Browser Use validates the final result against output_model_schema on the agent. v4 puts the schema on the extract() call that reads the data, and returns it typed:

Python
TypeScript

- class SearchResult(BaseModel):
-     title: str
-     url: str
- agent = Agent(task="...", llm=llm, output_model_schema=SearchResult)
- history = await agent.run()
- result = history.structured_output
+ class SearchResult(BaseModel):
+     title: str
+     url: str
+ result = await stagehand.extract(
+     "Extract the title and URL of the top result",
+     SearchResult,
+ )
+ # result.data is a SearchResult

+ const { data } = await stagehand.extract(
+   "Extract the title and URL of the top result",
+   z.object({ title: z.string(), url: z.url() }),
+ );

Calling extract() with no schema returns { extraction: string }. See extract.

Sensitive data

Both frameworks keep secrets out of the model’s context. Browser Use uses a sensitive_data dict of placeholder-to-value; v4 uses variables with %name% placeholders in the instruction, on act() and observe():

Python
TypeScript

- agent = Agent(
-     task="Log in with x_user and x_pass",
-     llm=llm,
-     sensitive_data={"x_user": "user@example.com", "x_pass": os.environ["PW"]},
- )
+ await stagehand.act(
+     "type %username% into the email field",
+     variables={"username": "user@example.com"},
+ )
+ await stagehand.act(
+     "type %password% into the password field",
+     variables={"password": os.environ["PW"]},
+ )
+ await stagehand.act("click the login button")

+ await stagehand.act("type %username% into the email field", {
+   variables: { username: "user@example.com" },
+ });
+ await stagehand.act("type %password% into the password field", {
+   variables: { password: process.env.PW },
+ });
+ await stagehand.act("click the login button");

Stagehand exposes only the variable names to the model and substitutes the real values locally. One exception: with server-side caching on, variable values travel to the cache service, so turn cache off for calls that carry credentials. See act. For Browser Use’s TOTP support (sensitive_data keys ending in bu_2fa_code), generate the code in your own script and pass it as a variable; v4 has no built-in 2FA step.

Custom tools

Browser Use’s @tools.action(...) / @controller.action(...) decorators register functions the agent can call. In v4 there’s no agent to register them with, so each custom action becomes an ordinary function you call directly in code mode, or one entry in your tool-calling tool set. The function body ports as-is; only the registration goes away.

Browser configuration

Browser(...) (aliased BrowserSession) options map onto the browser factory. The common ones:

Browser Use	Stagehand v4
`Browser(headless=False)`	`local_browser.launch({ headless: false })`
`Browser(cdp_url="http://localhost:9222")`	`local_browser.connect({ cdpUrl: "http://localhost:9222" })`
`Browser(proxy=ProxySettings(...))`	`proxy` on `local_browser.launch()`, or Browserbase proxies
`Browser(allowed_domains=[...])` / `prohibited_domains=[...]`	`browser.context.setDomainPolicy({ allowedDomains, blockedDomains })`
`Browser(storage_state="auth.json")`	Cookie API on `browser.context`, or a Browserbase context
`Browser(user_data_dir=...)`	`userDataDir` on `local_browser.launch()`
`Browser(accept_downloads=True, downloads_path=...)`	`acceptDownloads` and `downloadsPath` on `local_browser.launch()`
`Browser(keep_alive=True)`	`keepAlive` on `local_browser.launch()`
`@sandbox(...)` cloud deployment	`browserbase.launch({ apiKey })` plus deployments

Browser Use’s @sandbox decorator runs the agent next to the browser on cloud infrastructure. The v4 equivalent is a Browserbase browser: browserbase.launch() gives you the hosted session, and the live view, recording, and network detail replace @sandbox’s on_browser_created / live_url and on_log callbacks.

Quick reference

Browser Use	Stagehand v4
`Agent(task=..., llm=...)`, `agent.run()`	Code mode, or a tool-calling loop you own
The `task` string	Explicit steps: `page.locator()`, `act()`, `extract()`
`run(max_steps=...)`	Your loop bound, or the length of the script
`agent.add_new_task(...)` (follow-up)	Keep calling steps on the same browser
`ChatOpenAI`, `ChatAnthropic`, `ChatGoogle`	`model: { modelName: "provider/model", apiKey }`
`ChatBrowserUse()`	Omit `model` on a Browserbase browser (Model Gateway picks)
`page_extraction_llm`	`model` option on the `extract()` call
`output_model_schema=Model`	`extract(instruction, Model)`
`sensitive_data={...}`	`variables` on `act()` and `observe()`
`@tools.action(...)`, `@controller.action(...)`	A plain function, or one tool in your tool set
`tools` / `controller` registry	Your own tool set, see tool calling
`history.final_result()`	The return of your last `extract()`
`history.structured_output`	`extract().data`, typed by the schema
`history.urls()`, `history.errors()`	Your own bookkeeping between steps
`use_vision=True`	`screenshot: true` on an `extract()` call
`generate_gif=True`	Browserbase session recording
`Browser(...)` / `BrowserSession(...)`	`local_browser.launch()` or `browserbase.launch()`
`Browser(cdp_url=...)`	`local_browser.connect({ cdpUrl })`
`Browser(allowed_domains=...)`	`browser.context.setDomainPolicy({ allowedDomains })`
`Browser(storage_state=...)`	Cookie API, or a Browserbase context
`@sandbox(...)`	`browserbase.launch({ apiKey })`
`calculate_cost=True`, `history.usage`	`await stagehand.metrics()`

Troubleshooting

ImportError: cannot import name 'Agent'. There is no Agent in v4. Replace Agent(task=...).run() with code mode or tool calling. Nothing reads my API key. Stagehand reads no environment variables. Pass the Browserbase key to browserbase.launch() and any model key in the model option. load_dotenv() only populates os.environ; you still pass the values in. My script has no history object to read results from. There’s no run history. The value you’d have read from history.final_result() or history.structured_output is the return of your last extract() call, on .data. My custom @tools.action function has nowhere to register. Call it directly in code mode, or add it to your tool set for tool calling. Only the decorator goes away; the function body is unchanged. A retried step repeats a side effect. You’re retrying act(), the same hazard Browser Use’s max_failures loop had. Retry observe() instead and pass the resulting action to act() once. My generated script uses Agent or old Stagehand APIs. The assistant is drawing on Browser Use and older Stagehand patterns in its training data. Install the rule files from AI rules.

Next steps

AI rules

Set your coding assistant up to write v4 code

Act

Perform one action, or replay an observed one

Extract

Pull typed data, the home for structured output

Caching

Cut inference out of a stable flow

First steps

The basics

Configuration

Integrations

Best practices

Migration guide

SDK reference

Migrate Browser Use to v4

Why there’s no Agent

Hello world, side by side

Code mode

Tool calling

Let a coding assistant do the rest

Recommended migration order

Breaking changes

Initialization and teardown

The task string becomes explicit steps

Models

Structured output

Sensitive data

Custom tools

Browser configuration

Quick reference

Troubleshooting

Next steps

AI rules

Act

Extract

Caching

​Why there’s no Agent

​Hello world, side by side

​Code mode

​Tool calling

​Let a coding assistant do the rest

​Recommended migration order

​Breaking changes

​Initialization and teardown

​The task string becomes explicit steps

​Models

​Structured output

​Sensitive data

​Custom tools

​Browser configuration

​Quick reference

​Troubleshooting

​Next steps

AI rules

Act

Extract

Caching

Why there’s no Agent

Hello world, side by side

Code mode

Tool calling

Let a coding assistant do the rest

Recommended migration order

Breaking changes

Initialization and teardown

The task string becomes explicit steps

Models

Structured output

Sensitive data

Custom tools

Browser configuration

Quick reference

Troubleshooting

Next steps