convez.ai
SkillAutomaçãoIntermediário
superbereza

Browser Skill

superbereza/browser-skill

Automação de tarefas de navegador para agentes com snapshots interativos

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill Python que expõe um navegador real controlado por agentes de código através de um daemon CLI persistente com Playwright/Patchright. O agente interage com snapshots compactos da árvore de acessibilidade, clicando e digitando via refs curtas, com fallback para seletores CSS, coordenadas e JavaScript.

Instalação

$npx convez-ai superbereza/browser-skill

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 6 dias
Versão
v0.3.1
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automação de fluxos web que normalmente seriam manuais
  • Scraping e extração de dados de sites dinâmicos
  • Testes e validação de interfaces web em tempo real

Veredito

Ferramenta sólida para automação de browser via agentes, com design inteligente de refs + daemon persistente que reduz overhead de reconexão. Documento técnico completo e status estável (v0.2.0), mas captcha solver ainda está em scaffolding.

README

A tradução para o português deste README ainda está na fila.

browser

An adaptation of Playwright into a small CLI for driving a real browser from a coding agent. Same engine underneath — exposed as a handful of browser commands on a persistent daemon, with opinionated defaults chosen to make agent-driven browsing actually pleasant.

The agent sees each page as a compact snapshot of the accessibility tree, where every interactive element carries a short ref like [ref=e7]; it clicks and types by ref, with a fallback ladder down to CSS selectors, coordinates, and JS. A background daemon keeps one connection and the page's refs in memory, so commands share state without reconnecting — it starts on first use, stops when idle.

The browser is real and headed, on a dedicated profile you sign into once. Launch targets macOS; commands work cross-platform once attached.

Meant for automating browser tasks you'd otherwise do by hand.

How it works

Built on Patchright, a patched Playwright fork, attached to the browser over CDP. The patches keep the agent's automation traffic from tripping the CDP-level signals some sites use to flag automation, so ordinary use behaves like an ordinary session. Everything else is plain Playwright underneath.

The pieces:

  • CLI + daemon — a thin browser … client talks to a background daemon holding one Patchright/CDP connection, the current page, and the snapshot refs. State lives in the daemon; it auto-starts on first use and auto-stops when idle.
  • Snapshot — Playwright's aria_snapshot plus [ref=eN] refs; click/type by ref, falling back to CSS selectors, pixel coordinates, and JS.
  • Profile — a dedicated profile you sign into once; sessions persist across runs.
  • Captcha — detected deterministically; the agent works through it with the vision tools or hands it to you, with an optional solver you can wire up.

Install

This repo is its own marketplace:

/plugin marketplace add superbereza/browser-skill
/plugin install browser@browser-skill

The first browser call bootstraps a venv (Python >=3.10) and installs deps; browser launch starts a dedicated-profile browser and attaches.

macOS for launch. The dedicated profile is signed out on first run — log in once and sessions persist. Full usage: skills/browser/SKILL.md.

Design notes

Rationale for the non-obvious calls lives in docs/decisions/:

Status

v0.2.0. Smoke-tested on macOS: launch/attach, snapshot + refs, navigation, click/type/press/scroll/hover/select, tabs, upload/download, dialog, eval, wait. The optional playwright-captcha solver path is scaffolded, not yet wired. Per-site guides for airbnb / aviasales / chatgpt ship under skills/browser/site-guides/.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados