Autobrowse

Autobrowse provides a structured environment for teaching an agent how to complete a particular website task consistently. It executes the workflow, records what happened, and uses the evidence from each run to refine the task’s navigation strategy. The resulting knowledge can later be packaged as a standalone Claude Code skill.

Highlights

• Runs configurable evaluation cycles for individual tasks or groups of tasks
• Captures summaries, action histories, costs, screenshots, and failure evidence
• Supports local browsers and remote Browserbase sessions for more demanding sites
• Can combine agent activity with browser-level network and console diagnostics
• Generates optional Playwright or Stagehand implementations from successful runs

Autobrowse is aimed at agent engineers, automation developers, and teams that need repeatable browser behavior instead of fragile one-off prompts. It requires Node.js 18 or newer, the browse CLI, and an Anthropic API key; some remote and tracing features need additional Browserbase credentials or a companion skill. Install it from the Autobrowse folder in Browserbase’s public skills repository.