yarin-zhang Browser Task Router
Why this exists
This project prefers ego-browser (ego-lite) for browser automation. It runs in its own isolated space and reuses the user's existing login state, so it doesn't compete with or disrupt the user's own browser session the way driving their real Chrome would. The built-in Claude_Browser and claude-in-chrome tools are general-purpose fallbacks — useful, but not the default here.
What to do
When a task needs browser automation — opening a page, filling a form, clicking, taking a screenshot, scraping/extracting page data, testing a web app, or logging into a site:
- Invoke the
Skilltool withskill: "ego-browser", passing the task details throughargsso it has full context (target URL, form values, what to extract, etc.). - Let ego-browser's own instructions drive the actual browser interaction rather than duplicating its logic here.
- Don't call
mcp__Claude_Browser__*ormcp__claude-in-chrome__*tools directly for this kind of task unless one of the exceptions below applies.
When not to route to ego-browser
- The user names a different tool or browser explicitly — "open it in my real Chrome", "use claude-in-chrome", "use the preview browser". Honor that instead; an explicit ask overrides the default.
- The task is previewing your own local dev server or build output, not automating an external site. That's what
Claude_Browser'spreview_startis for — it's about verifying the app you're building, a different job from the external-website automation ego-browser targets. - ego-browser is unavailable or errors out. Fall back to the built-in browser tools and tell the user why you fell back.
How to invoke
Skill({ skill: "ego-browser", args: "<the user's original browser task, with details intact>" })
Pass through enough of the original request that ego-browser doesn't have to guess — don't compress away URLs, field values, or extraction targets while summarizing.