← ClaudeAtlas

agentic-browserlisted

How to use the shared agent browser (the `browser` MCP server, tools named browser_*, from agentic-playwright-mcp). Read before the first browser call in a chat, and before spawning subagents that will use the browser. Covers tab groups, subagents, logins and credentials, tab links, what is shared between chats, and the rule to report every browser problem to the user.
UserNamedMartin/agentic-playwright-mcp · ★ 0 · AI & Automation · score 70
Install: claude install-skill UserNamedMartin/agentic-playwright-mcp
# Shared agent browser The `browser` MCP server is a real, normal Chrome that runs minimized in the background and is shared by every agent of this setup. Logins live in it permanently. It is Playwright MCP underneath, so the usual `browser_*` tools work, with a few differences described here. ## Your tabs - Your chat gets its own tab group, titled after the chat. You only see and control tabs you opened (plus pages your tabs open). Other chats cannot see yours and you cannot see theirs. - Just call `browser_navigate`; the first call opens your first tab. More tabs: `browser_tabs` with action `new`. `browser_tabs` lists your tabs with short ids; pass `"tab": "<id>"` to any tool to act on a specific tab. - Links that open a new tab open it in the background inside your group; they do not switch your current tab. Check `browser_tabs` after clicking one. - Close tabs you no longer need with `browser_tabs` action `close`. Do not use `browser_close`. - Everything happens in the background: the user does not see the window unless you give them a link (see "Showing pages to the user"). Never assume they are watching. - A chat forked in the Claude desktop app starts with copies of the original chat's tabs (listed under "Tabs of the original chat" in your first result); the originals stay with that chat. The copies are reloaded pages: take a snapshot before acting, and on payment, one-time or form-result pages check that the copy did not repeat or break some