Playwright CLI
The video demonstrates three browser automation use cases using Playwright CLI connected to Claude Code: automated QA testing and bug fixing for a web app, scraping dentist contact information from the web, and automating interactions on a platform requiring login. The presenter also showcases a fully autonomous community management bot built on these techniques.
Summary
The presenter walks through practical browser automation use cases by connecting Playwright CLI to Claude Code, explaining upfront that he chose Playwright CLI over the Chrome DevTools MCP server primarily because MCP's extensive tool descriptions consume too many tokens.
The first use case involves building a multi-page onboarding form (12 questions) and then having Claude Code autonomously QA it using a headed browser. Without manual intervention, the bot fills in form fields, takes screenshots, identifies bugs (such as the Enter key not advancing past the textarea page and the review page failing to load), fixes them in the source code, and reruns the tests until they pass. The presenter frames this as a transformative capability for software development, where multiple headless browser instances could continuously test, find bugs, and self-correct.
The second use case is web scraping: building a Playwright script to find dentist offices in California and collect their phone numbers. When Google blocked the automation, Claude Code adapted by switching to DuckDuckGo. The presenter stepped away briefly and returned to find five dental office phone numbers already collected, highlighting how the script self-corrects during execution.
The third use case addresses whether browser automations can work in authenticated sessions. Claude Code uses a persistent browser profile linked to the user's existing Chrome data, allowing it to access a Skool community already logged in. The task is to navigate to the 'wins' channel and like posts. Initial attempts resulted in the bot repeatedly toggling likes on and off. After a few guided iterations — including instructions to filter by newest posts and identify liked posts by their yellow thumbs-up icon — the bot successfully liked posts while skipping already-liked ones.
In an extended segment, the presenter reveals a fully operational autonomous agent ('AIS agent') running in his Skool community, posting daily AI news roundups, responding to notifications, and even autonomously making a birthday post. These automations run on a schedule via the Claude Code desktop app in headless mode. The agent continuously builds new scripts as it encounters new UI challenges, such as learning how to vote on polls. The presenter closes by comparing Playwright CLI to other browser automation CLIs, noting he values token efficiency and script-learning performance over speed.
Key Insights
- The presenter chose Playwright CLI over the Chrome DevTools MCP server specifically because the MCP server's large number of tools and their descriptions consume an excessive number of tokens, making it less efficient for browser automation tasks.
- During the Google dentist search, Claude Code autonomously detected that Google was blocking the automation and switched to DuckDuckGo without any instruction from the presenter, successfully collecting five dental office phone numbers.
- The automated QA bot not only identified specific bugs — including the Enter key failing to advance past the textarea and the review page never loading — but also fixed them in the source code and reran tests autonomously until they passed, without the presenter intervening.
- The presenter's autonomous Skool community agent ('AIS agent') independently made a birthday post for him based on information it already knew, without being explicitly instructed to do so, demonstrating emergent agentic behavior beyond its programmed tasks.
- The presenter argues that the key performance metric for browser automation CLIs is not speed but a balance of token efficiency and script-learning ability — specifically how well the tool improves its clicking and navigation accuracy across repeated runs.
Topics
Full transcript available for MurmurCast members
Sign Up to Access