Webwright

Why we included this project

Webwright takes a deliberately different route from most browser agents. Instead of predicting one click or keystroke at a time inside a persistent session, it hands the model a terminal and lets it write short Python scripts that drive Playwright browser sessions as needed. Because the actual work lives in a runnable script in your workspace rather than in browser state, multi-step tasks like filling forms across several pages are much easier to repeat and debug. The core loop is a single agent file of roughly 450 lines on top of Playwright, so it is readable end to end and easy to fork, and it works with OpenAI, Anthropic, or OpenRouter models. A skill-factory mode distills solved tasks into parameterized scripts that rerun without a model, and the same skills load into Claude Code, Codex, or OpenClaw through included plugin manifests.