browser-use: how to connect AI agents to the browser

If you have tried to extract data from a site with no public API or automate a flow that requires login, browser-use solves that case. It combines a language model with browser automation via Playwright. The model decides the next action and the library executes it in the browser.

The browser-use/browser-use repository is one of the most active on GitHub in the agents space. It has more than 120 thousand stars, MIT license and was updated on November 9, 2026. That gives confidence to adopt it in real projects.

In this article you will see:

• What browser-use does and how it controls the browser
• How to install and run your first agent
• Real use cases for devs and founders
• Limits, costs and security considerations
• How to integrate with an agent squad

What is browser-use and how automation works

A traditional code agent works with files, terminals and APIs. When the task involves a site with no API, it stalls. browser-use closes that gap. It opens a controlled browser, captures the page state and asks the model which action to take next.

The cycle is simple. The agent receives a goal in natural language. It observes the current page, chooses an action, such as clicking a button or filling a field, and repeats until done. The library handles execution via Playwright, which is stable and supports Chrome, Firefox and WebKit.

Unlike traditional scrapers based on fixed selectors, browser-use adapts when the layout changes. The model understands the page by its content and structure, not by a fragile selector. This reduces maintenance, but does not eliminate the need for tests.

How to install browser-use and run your first agent

Installation requires Python and Playwright. browser-use handles most of the setup.

$ pip install browser-use
$ playwright install chromium

Then configure the model key. browser-use supports OpenAI, Anthropic, Groq and other providers compatible with the OpenAI API.

from browser_use import Agent
from langchain_openai import ChatOpenAI
import asyncio

async def main():
    agent = Agent(
        task="Abra o GitHub e pesquise por 'mem0ai/mem0'.",
        llm=ChatOpenAI(model="gpt-4o"),
    )
    await agent.run()

asyncio.run(main())

The run method executes the cycle until the agent considers the task complete or hits the step limit. You can set the maximum number of steps, the timeout and whether the browser should run in visible or headless mode.

If you prefer not to use OpenAI, point the llm to another compatible provider. The official documentation includes examples for Anthropic, Groq and local models via Ollama.

Real use cases for devs and founders

browser-use shines in three situations. The first is data extraction from sites with no API, such as supplier portals, admin dashboards and pricing pages. The second is automation of flows that require human interaction, such as filling multi-step forms or downloading reports. The third is interface testing guided by natural language.

For founders, a common case is monitoring competitors. The agent can open a competitor's pricing page, capture the plans and save them to a file. For devs, automating account creation in test environments saves hours.

If you already use agents for code, browser-use complements them. One agent can write the code and another can open the browser to test the result. In multi-agent flows, each agent in its own isolated space avoids conflicts.

Limits, costs and security considerations

Browser automation has costs that do not show up in the code. Each step consumes model tokens, because the page needs to be read and interpreted. Long tasks can consume a lot. Set clear limits for steps and time.

Sites with CAPTCHA, strong authentication or bot detection can block the agent. browser-use does not solve CAPTCHA by default. In corporate environments, check the site's terms of use before automating.

Security is another point. An agent with access to a browser can perform irreversible actions, such as submitting forms or deleting data. Always run in an isolated environment, with test credentials, and require human confirmation for critical actions.

How to integrate browser-use with an agent squad

browser-use exposes a simple interface that can be called by any orchestrator. You can create one agent specialized in navigation and another in analyzing the extracted data. Each in its role, with its own models and keys.

If you coordinate multiple agents, it is worth separating navigation memory from analysis memory. An agent that remembers the selectors that worked saves time in future runs. Also see How to run agents in parallel in VS Code and Multi-agent workflows on bridgecode to keep each one in its isolated space.

Frequently asked questions about browser-use

Is browser-use free?

The library is open source with an MIT license. The cost comes from the language model you choose and the infrastructure where the browser runs.

Do I need a GPU to run browser-use?

No. The browser runs on CPU. The language model can be remote. If you use a local model, then a GPU may be needed for that model.

Does browser-use solve CAPTCHA?

No by default. Sites with CAPTCHA or bot detection can block the agent. There are third-party services for this, but check the site's terms of use.

Can I use browser-use in production?

Yes, as long as you handle security, costs and failures. Run in an isolated environment, set step limits and monitor the results.

Does browser-use work with Firefox and Safari?

It uses Playwright, which supports Chromium, Firefox and WebKit. WebKit support covers the Safari engine. Check the Playwright version in the documentation.