Browser Automation

What's in this pack

This pack collects the seven browser automation assets that come up over and over again when serious agent setups are reverse-engineered from public configs and Twitter screenshots. Three are the headline open-source frameworks. Three are scraping templates that wrap them. One is the operational MCP fix that keeps Chrome from breaking when multiple agents share the same browser.

Why this pack exists

Most "AI scraper" tutorials stop at "open this page and read the text." Real browser automation has to deal with login walls, cookie banners, modal popups, infinite scroll, and the fact that multiple agents trying to drive the same Chrome session will fight each other. This pack picks the seven assets that solve those problems specifically — not the 200 starred-but-abandoned alternatives.

The three headline frameworks each have a niche:

• Browser-Use is the easiest entry point. Python-native, ships with a Playwright wrapper that an LLM can talk to in plain English. Best for one-off automations and prototypes.
• Skyvern runs as a server and uses vision + LLM together. Best when the page changes structure between visits and selectors break. It's slower than Browser-Use but more resilient.
• Stagehand is the production option. TypeScript-first, made by the Browserbase team, designed to run thousands of agent-driven sessions. If you're shipping a customer-facing product, this is the one.

The Chrome MCP popup fix is included because every multi-agent Chrome setup eventually hits the bug where one agent's modal blocks another agent's click. The included MCP server intercepts the most common popup patterns and clears them automatically.

Install in one command

# Install the entire pack
tokrepo install pack/browser-automation

# Or pick the framework you actually need
tokrepo install browser-use
tokrepo install skyvern
tokrepo install stagehand

The TokRepo CLI installs each asset into the correct location for your AI tool — Claude Code subagents into .claude/agents/, Cursor rules into .cursor/rules/, AGENTS.md entries for Codex CLI, and so on. Pull whichever subset matches your stack.

Watch out for

• Don't run Browser-Use and Stagehand on the same Chrome profile simultaneously. They'll race for the active tab. Use separate user-data-dirs or a remote Browserbase session.
• Skyvern needs a GPU or a cheap vision model fallback. The default config calls GPT-4o for every screen — that adds up fast on long-running scrapes.
• Login-walled scraping breaks reCAPTCHA Terms of Service. Stick to your own dashboards or sites that explicitly allow automation. The pack ships with a "session export from real browser" pattern so you don't need to log in via the agent.
• Cookie banners are not all GDPR. Many sites use them as anti-bot tarpits. The popup-fix MCP solves the common ones; the rest you'll have to add manually.
• Headless detection is real. Sites like Cloudflare and DataDome will fingerprint you. If you need to bypass, use Stagehand with Browserbase — they handle the stealth layer.

When this pack alone isn't enough

If your goal is large-scale crawling of a single high-value site (job boards, real estate listings, e-commerce), a dedicated scraper like Apify or a custom Playwright cluster will outperform agent-driven automation by 5-10x in cost. Browser automation agents are best when:

• You don't know in advance which fields you need
• The page layout varies between visits
• You want to glue scraping into a larger agent task (research, lead gen, QA)

For pure throughput, drop the LLM and write the selectors yourself. For flexibility and the ability to recover from layout changes, this pack wins.