Browser Automation as an Installable Agent Skill Pack

·https://github.com/browserbase/skills
repoaiagentsbrowser-automationcliclaude-codebrowserbaseskillsui-testing

Maps to joelclaw's skill system: installable browser skills plus traces make agent browser work repeatable, reviewable, and less vibes-based.

Browserbase Skills is a plugin-style skill pack for Claude Code that wraps Browserbase browser automation and the official browse CLI. The interesting bit isn’t “agent can click browser.” We already know that trick. The interesting bit is browser work packaged as named, installable capabilities.

The repo splits browser work into separate skills: browser for automation, browser-trace for CDP traces, browser-to-api for turning observed traffic into OpenAPI 3.1, autobrowse for self-improving automation, safe-browser for domain-gated agents, and agent-experience for testing how agent-friendly a product or docs site is.

That shape is useful for joelclaw because it treats skills as operational affordances, not prompt seasoning. A browser trace becomes evidence. A local cookie sync becomes an auth bridge. A UI test becomes a reproducible review step instead of “the agent poked around and said it looked fine,” which is the kind of bullshit that makes agent work hard to trust.

The sharp pattern here is the bundle: browser automation, trace capture, safe boundaries, API discovery, and adversarial UI testing all living beside each other as skills. That maps cleanly to agent loops where the implementor can change code, the reviewer can inspect browser evidence, and the system can keep the receipt instead of relying on screenshots floating around in chat.

Key Ideas

  • Browserbase Skills packages browser automation for Claude Code as installable skills, not one giant browser prompt.
  • The browser-trace skill captures CDP events, screenshots, and DOM dumps, which turns browser automation into inspectable evidence.
  • The browser-to-api skill uses observed HTTP traffic to draft an OpenAPI 3.1 spec, which is a neat bridge from messy websites to structured contracts.
  • The autobrowse skill closes the loop by running a browsing task, reading the trace, and improving the navigation skill until it passes.
  • The safe-browser skill puts a domain allowlist around browser agents, which matters when authenticated browser state is involved.
  • The cookie-sync skill treats local Chrome cookies as a bridge into a Browserbase persistent context.
  • The ui-test and agent-experience skills turn browser automation into adversarial product QA and agent UX auditing.

Links

← All discoveries