Browser Skill
superbereza/browser-skill
Automação de tarefas de navegador para agentes com snapshots interativos
Ver no GitHubSkill 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-skill1 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 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_snapshotplus[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/:
- Patchright as the engine
- Persistent daemon, not a stateless CLI
- Built-in
aria_snapshot, no custom AX filter - Dedicated profile
- Captcha handling
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.