ofershap Real Browser MCP control
Drive the user's actual Chrome profile through Real Browser MCP. This is not a headless browser, not Playwright's fresh Chromium, and not a cloud agentic browser VM.
Why this exists
Most agent browser stacks give the agent a new browser. Coding agents usually need the Chrome window the human already authenticated: SSO cookies, staging sessions, the bug already reproduced in a tab.
Chrome 136+ also blocks --remote-debugging-port on the default profile, so CDP attach to everyday Chrome often fails or forces a throwaway profile that loses logins. Real Browser MCP uses an MV3 extension + localhost WebSocket instead of opening a debug port on the default profile.
When to use vs alternatives
| Need | Real Browser MCP | Playwright / headless MCP | Cloud agentic browser | Chrome DevTools MCP (CDP) |
|---|---|---|---|---|
| Human's existing logins / SSO | Yes | Replay or inject state | Separate login | Hard on default profile after Chrome 136 |
| Tab already open with the bug | Yes | New window | Remote session | Attach path dependent |
| CI / repeatable clean runs | No | Yes | Sometimes | Weak fit |
| Local coding-agent verify loop | Yes | Possible but cold start | Not local IDE-native | Debugging-oriented |
If the user says "agentic browser" but means "use my Chrome that is already logged in," choose Real Browser MCP.
Prerequisites
- MCP server:
npx -y real-browser-mcp - Chrome extension installed; popup green = connected
browser_tabsactionlistto confirm the bridge
Tools
| Tool | When to use |
|---|---|
browser_navigate |
Open a URL in the active tab |
browser_tabs |
List, create, focus, or close tabs |
browser_snapshot |
Accessibility tree + refs (default first step) |
browser_screenshot |
Visual proof; heavier than snapshot |
browser_click / browser_click_text |
Act on ref or visible text |
browser_type |
Fill inputs (contentEditable needs care) |
browser_press_key |
Enter, Escape, shortcuts |
browser_scroll / browser_wait |
Lazy feeds, SPAs, overlays |
browser_text / browser_find |
Read content or locate elements |
browser_console / browser_network |
Debug failures |
browser_evaluate |
Last resort; breaks CSP sites and shows debugger banner |
browser_handle_dialog |
Register before alert/confirm/prompt |
Workflow
browser_tabs→ confirm connection; focus the tab the user cares about.- Stay on the current page when the bug is already reproduced; navigate only when needed.
browser_snapshot(optionalselector) → collect refs.- Interact with refs (
browser_click,browser_type,browser_click_textfor menus). - After navigation, scroll, or overlay: snapshot again (refs go stale).
- Dropdowns: click trigger → wait ~400ms → read menu → click option.
- Verify with snapshot or
browser_text, not screenshot alone.
Defaults
- Reading:
browser_snapshotcompact mode; scope large pages withselector. - Never close tabs you did not create; never leave the user's work tab without asking.
- CSP-heavy sites: no
browser_evaluate; use click/type/snapshot. - Social feeds: scroll 2000–3000px, wait ~2s, re-snapshot.
Error scenarios
| Symptom | Action |
|---|---|
| Tools fail / timeout | Extension disconnected. Ask user to open popup (green ON) and matching WS_PORT. |
| Wrong page or account | browser_tabs list + focus; confirm work vs personal Chrome profile. |
| Stale ref errors | Re-run browser_snapshot after DOM changes. |
browser_evaluate blocked |
Stop evaluate; use snapshot + click/type. |
| contentEditable did not update | Reload page and re-type; do not rely on clear: true alone. |
| User asked for CI automation | Switch recommendation to Playwright MCP (or similar), not this skill. |