Webwright

Normal
GitHub Python MIT

Description

A simple SWE style browser agent framework that achieves SOTA results on long horizon web tasks.

Key Features

  • Coding agent with terminal access that writes and executes Playwright scripts instead of predicting individual browser actions
  • Workspace-as-state architecture where code and logs persist, not browser sessions
  • Lightweight core at ~1.5k LoC with pluggable model backends (OpenAI, Anthropic, OpenRouter)
  • SOTA performance: 86.7% on Online-Mind2Web and 60.1% on Odysseys long-horizon tasks
  • Task2UI mode that renders task results into reusable HTML-based web applications
  • Seamless integration as a Claude Code skill, Codex plugin, or standalone CLI tool

Use Cases

πŸ’‘ Complex web automation tasks requiring multi-step workflows like form filling and date selection
πŸ’‘ Long-horizon research tasks spanning dozens of pages with iterative exploration
πŸ’‘ RPA script development where browser sessions are disposable and code is reusable
πŸ’‘ Web testing and QA across dynamic sites with lazy loading and re-rendering
πŸ’‘ Converting natural language web tasks into executable Playwright scripts

Tags

browser agent automation python web-agent

Categories

🌐 Browser Agent πŸ€– Agent Framework

Quick Start

Install via pip (pip install webwright) and set up with playwright install. Run with: webwright --task "your task description" --backend openai. The agent writes Python scripts, executes them in browser sessions, inspects screenshots, and iterates until the task is complete.
Visit GitHub

Related Projects