AutoBrowse: A Karpathy-Inspired Skill That Lets Claude Code Control the Browser

A new Claude Code skill called AutoBrowse takes inspiration from Andrej Karpathy's auto-research harness. It gives your coding agent the ability to perform tasks on the web by exploring pages through a browser-based CLI, learning from failed attempts, and iterating until it lands on a reliable workflow.

How It Works

AutoBrowse connects Claude Code to a headless browser. You give the agent a task, and it explores the target page, reads the DOM, and tries to accomplish what you asked. When something goes wrong, it logs the failure, adjusts its approach, and tries again. Each iteration brings it closer to a repeatable workflow.

The demo shows the skill running against Google Flights, searching for round-trip economy tickets. The agent navigates the calendar, selects dates, and extracts structured flight data including airline, price, duration, stops, and times. The output is clean JSON with fields like airline, price, departure_time, and arrival_time.

At the end of a successful run, the skill prompts you to save a SKILL.md file so Claude can reuse the learned workflow in future sessions. This is the convergence step: once the agent finds a path that works, it graduates into a repeatable skill.

Vet Before You Trust

The creator emphasizes that you need to vet skills like this before trusting them. Browser automation skills have broad access by nature, and running one without reviewing its code first is a risk worth taking seriously.

Key Takeaways

• AutoBrowse lets Claude Code perform web tasks by controlling a browser, inspired by Karpathy's auto-research approach
• The skill iterates on failures automatically, converging on a reliable workflow over multiple attempts
• Always review and vet third-party skills before running them in your environment

Resources

• AutoBrowse Skill -- The skill page on skills.sh
• LvcidPsyche/auto-browser -- GitHub repo for the underlying auto-browser project