browser-use

Sweet spot: a developer who already works in Python and has a specific web-automation job to do. Drop browser-use into a script, point an LLM at it, and you will have a working agent in an afternoon. It is the most honest middle ground between headless scraping (brittle) and full-blown commercial RPA (expensive). Failure modes to know about. First, cost creeps: every page snapshot is tokens, so a 50-step agent on GPT-4o can hit a few dollars per run — cache aggressively, use smaller models for routine steps. Second, model capability is a hard floor — if your agent is deciding between visually similar buttons, you need vision-capable models. Third, anti-bot pages still win: browser-use is a tool, not a stealth solution. What to pilot before committing. Build one agent against one flow. Measure success rate over 20 runs, cost per run, and wall-time per run. If success rate is above 90% on your target flow, it will scale; below 70% and the prompt/model needs work before you invest more.

browser-use is an open-source Python framework that turns any LLM into a web agent. It wraps Playwright with a DOM-aware action layer — the model receives an annotated snapshot of the page and decides whether to click, type, scroll, or wait. Unlike screenshot-only approaches, browser-use feeds a structured element tree plus optional vision, which keeps token costs manageable on long flows. It is used as the scaffolding for a rapidly growing set of agent products: form-fillers, data-extraction bots, QA automations, e-commerce checkout agents, and research assistants. The library ships with memory, history recording, replay, and a built-in evaluator to run your agent against a scripted task set. The team also operates a hosted cloud (browser-use.com) that provides remote browser sessions, concurrency, session persistence, and an API key interface for teams that do not want to manage Playwright infrastructure themselves. The cloud tier is optional — everything the library does runs locally. It is one of the top 3 Python repos of 2024–2025 by both stars and real-world adoption, widely cited in agent tutorials and referenced by frameworks like LangChain and CrewAI as the recommended browser-automation primitive.