browser-use
Rong-Tao/browser-use
Automate seu Chrome autenticado via CLI, mantendo cookies e sessões intactas
Ver no GitHubSkill que conecta Claude a uma instância Chrome já autenticada via Chrome DevTools Protocol, permitindo automação de navegador com cliques humanizados, preenchimento de formulários, leitura de DOM e screenshots. Reutiliza sessões e cookies reais, eliminando necessidade de re-autenticação ou detecção de bot.
Instalação
npx convez-ai Rong-Tao/browser-use1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automação de tarefas web mantendo login persistente
- Scraping de dados em sites com autenticação
- Preenchimento de formulários e submissão de dados
- Testes automatizados com perfil real do usuário
- Captura de screenshots e análise de conteúdo dinâmico
Veredito

Ferramenta pragmática que resolve o problema real de automação web sem bot-detection, diferenciando-se por reutilizar sessões autenticadas. Implementação limpa em Python com CDP nativo, ideal para quem precisa de automação robusta sem headless profiles descartáveis.
README
A tradução para o português deste README ainda está na fila.
browser-use — drive your real, logged-in Chrome from the CLI
A small, site-agnostic command-line tool that attaches to your existing, already-logged-in Chrome over the Chrome DevTools Protocol (CDP, port 9222) and lets you (or an LLM agent) read pages, fill forms, click, upload files, and screenshot — reusing your real browser profile, so every site you're signed into stays signed in. No passwords, no re-auth, no fresh headless session that trips bot detection.
It ships as both a plain CLI and a Claude Code
skill (see SKILL.md) — the agent-facing operating doctrine lives
there.
⚠️ The name
browser-useis already taken on PyPI/GitHub by a popular, unrelated project. Pick a distinct name before publishing.
Why
Most browser automation spins up a fresh, headless, profile-less Chrome — which means re-logging-in everywhere, fighting 2FA, and tripping bot detection. This tool does the opposite: it attaches to the Chrome you already use, with your cookies and sessions intact, and drives it the way a person would.
- Reuses your real login — cookies, sessions, 2FA all carry over.
- Human-like clicks by default — the pointer glides to the target along a curved, eased, slightly-irregular path (acceleration/deceleration, jitter, different every time), so sites that score pointer movement see a human.
- Read-first discipline — every command reads state back from the DOM; you're told to confirm "it worked" from the page, never from an exit code.
- Works against a remote Chrome too — run the CLI on the same machine as the
Chrome it drives (CDP is
localhost:9222); drive that machine over SSH.
Install (uv)
Requires uv and Google Chrome.
cd scripts
uv sync # create the venv, install deps
uv run playwright install chromium # first run only, if Playwright asks
Quick start
cd scripts
uv run python main.py open-chrome # launch Chrome with the debug port
uv run python main.py open https://github.com # navigate (reuses the active tab)
uv run python main.py get-tab-info # read the page (text) — start here
uv run python main.py click "Sign in" # human-like click by visible text
uv run python main.py screenshot # capture a PNG, then view it
Command reference
| Command | What it does |
|---|---|
open-chrome [-p PROFILE] / chrome-status | launch / check the debug Chrome |
profiles | list Chrome profiles (accounts) |
open <url> [--new-tab] | navigate (reuses active tab by default) |
goto <id-or-url> | jump to a known tab or URL |
get-tab-info [--mode text|links|inputs|html|selector] [-s SEL] | read the page; html strips base64 data: URIs |
screenshot [-s SEL] [--full-page] [-o OUT] | capture a PNG to view |
cookies [-d DOMAIN] | list cookies (names/expiry, not values) — confirm a login persisted |
box SELECTOR | an element's on-screen box (x/y/w/h + centre) |
fill SELECTOR VALUE [--enter] | type into an <input> (--enter for React/Ant tag-inputs) |
type TEXT [-s SEL] [--enter] | real keystrokes into a contenteditable rich editor; \n = Enter |
upload PATH [-s SEL] | set a file on a file <input> (even hidden ones) |
click TEXT… [-s SEL] [-n N] [--bias-x F] [--bias-y F] | human-like click by text or selector; --bias-x/-y aim at a fraction of the box (for a wide host wrapping several controls) |
click-xy X Y | human click at absolute coordinates (closed-shadow-DOM last resort) |
tabs [--all] / get-current-tab | list session tabs (or every open tab) / show current |
focus MATCH | switch active tab to an open one by URL/title substring |
Design philosophy (kept deliberately)
The tool is generic primitives only. Site-specific knowledge — how to log in to a particular site, how to publish a post there — belongs in a separate SOP that calls these primitives, never in the tool itself.
Two ideas from SKILL.md drive how an agent should use it:
- Explore vs. re-use. The first time through any flow is exploration: strict read→act→read, one action at a time, never batching blind clicks (which misfire and trip bot detection). Record the working steps as an SOP. Once an SOP exists, re-use it.
- Self-evolution. When a command can't do what a step needs, or a read view
buries what you need, improve the tool (a new generic command, a better
read mode) rather than working around it inline — then update
SKILL.mdso the next run inherits it. This repo grew exactly that way.
Architecture
scripts/
├── main.py # Typer CLI — thin command layer
└── browser/
├── chrome.py # launch Chrome w/ debug port, CDP connection, autodetect
├── page.py # the generic page primitives (read/fill/click/upload/…)
└── tabs.py # short-ID ↔ URL registry, fcntl-locked (safe across agents)
- The CLI never closes your Chrome — it detaches the CDP session (
p.stop());browser.close()would kill your windows. - Tab IDs persist in
scripts/.current_tab.json, cleared on eachopen-chrome, so a chain of commands keeps operating on the same tab.
License
MIT © 2026 Rong