Several agents, one Chrome

The problem this solves

The moment you run more than one agent against a real browser, they start fighting over it: one navigates away from a page another was mid-task on, tabs pile up with no owner, and you can't tell which session opened what. A single shared browser with no isolation turns parallel work into a mess.

How isolation works

Browser MCP scopes each session to its own set of tabs, visually separated into a color-coded tab group (Chrome's native tab groups, driven by the extension). When a session opens a tab, it lands in that session's group. When a session lists or switches tabs (browser_list_tabs, browser_get_new_tab, browser_switch_tab), it only sees its own - not the tabs another session is working in. You can glance at Chrome and read, by colour, which group belongs to which agent.

A concrete parallel setup

• Session A - research: navigating docs and dashboards, pulling content with browser_get_page_content. Blue group.
• Session B - testing: driving a form flow with browser_fill / browser_click, checking browser_console_logs. Green group.
• Session C - monitoring: watching a status page, taking periodic browser_screenshots. Yellow group.

All three run in the same Chrome, at the same time, on your real logged-in profile. None of them can accidentally act on another's tabs.

What it looks like

You:     What is open in this session?

Claude:  [browser_list_tabs]
         Two tabs, both in the blue group:
           #312  Analytics - Last 30 days
           #313  admin.example.com/orders

You:     And the tab the other agent is using?

Claude:  It is not in this list. browser_list_tabs only returns the tabs this
         session claimed - the other agent's tabs are in its own colour group,
         and I cannot see or touch them.

That is the whole model in one exchange: what you get back is your session, and only your session.

Why this matters for real-Chrome automation

Headless tools solve parallelism by spawning many isolated browsers - clean, but none of them are your logged-in browser. Browser MCP's whole point is to act as you, in your real Chrome. Multi-session tab-group isolation is what makes that safe to do more than one agent at a time: you keep the real-browser advantage without the sessions colliding.

FAQ

How many sessions can run at once? Up to 20 concurrent sessions, each in its own color-coded tab group (as documented in the project README).

Can one session see another session's tabs? No - that's the point of the isolation. Each session's tab operations are scoped to its own group.

Do the sessions share my login state? Yes - they all run in your real Chrome, so they share your cookies and sessions. Isolation is about tabs, not identity: every session acts as you.

How do I start a second session? Each MCP client session that connects gets its own group automatically. Point a second Claude Code (or Cursor / VS Code) session at Browser MCP and it lands in its own colour.

How do I set it up? Install Browser MCP, then run more than one agent session against it - the tab groups appear on their own.