Webwright — браузерный агент, пишущий код

★ 6K

Webwright is an open-source library from Microsoft that turns coding models — Claude Code, Codex, OpenClaw, Hermes — into browser agents: the LLM gets a terminal, launches Playwright browser sessions to inspect pages, and encodes every web task as a reusable, re-runnable Python script. Reach for it when you need to automate complex, long-horizon browser workflows — filling multi-step forms, scraping dynamic SPAs, navigating dynamic sites — and want a reproducible code file as the artifact rather than a fragile click history. It also fits when heavyweight multi-agent orchestration has become a maintenance burden and you want a minimal, fully transparent loop instead. The agent core is ~450 lines of Python, the Playwright environment ~570 lines; the only dependencies are httpx, pydantic, playwright, and typer. Backends supported: OpenAI, Anthropic, OpenRouter. Connects to Claude Code and Codex as a plugin with `/plugin install webwright@webwright`. The Skill Factory feature distills solved tasks into parameterized scripts that re-run standalone with no model in ~40 s and zero tokens. Task2UI mode renders task results as an HTML app. The direction is code-artifact generation from web sessions, not step-by-step browser control — for classic step-by-step RPA, browser-use or Stagehand are closer alternatives.