@ghostystudio/browser-mcp

MCP server that lends your own Chrome — with the sessions you already have open — to any coding agent (Claude Code, Cursor, Codex, any MCP client), through the Ghosty for Chrome extension. Same idea as Claude in Chrome: the agent reads pages (accessibility snapshot with refs), clicks, types, fills forms, uploads files, takes screenshots and works in several tabs at once. You see a lilac border, a named cursor and a ■ Stop button. It never types passwords.

Setup

1. Install Ghosty for Chrome: https://www.ghosty.studio/chrome

2. Local bridge (same computer, no internet, no token):

   npx -y @ghostystudio/browser-mcp install-host

   Reload the extension (chrome://extensions → ↻). Its panel says Terminal: connected.

3. Add it to your agent:

   claude mcp add ghosty-browser -- npx -y @ghostystudio/browser-mcp      # Claude Code
   codex mcp add ghosty-browser -- npx -y @ghostystudio/browser-mcp       # Codex

   Cursor (~/.cursor/mcp.json):

   { "mcpServers": { "ghosty-browser": { "command": "npx", "args": ["-y", "@ghostystudio/browser-mcp"] } } }

Register it once. Your agent starts the server once per session and keeps it alive, so each tool call costs milliseconds. Don't spawn npx per call (1–2 s each). For scripts, install it globally and use the call mode:

npm i -g @ghostystudio/browser-mcp
ghosty-browser-mcp call read_page '{"tabId": 123}'

Slow tools (over 3 s) send MCP notifications/progress ("click: waiting for the page to react… 5 s") when the client passes a progressToken; the extension panel shows the same.

Background tabs work: the extension keeps them alive (focus emulation + active lifecycle via CDP) and captures with Page.captureScreenshot, so it can read, click, type and screenshot tabs you are not looking at.

Remote (Chrome on another machine): set GS_BROWSER_TOKEN to the bt_… token from the extension panel («Connect the terminal») and the calls go through Ghosty Studio.

Tools

browser_status, browser_navigate, browser_read_page, browser_find, browser_click, browser_type, browser_fill_form, browser_select_option, browser_press_key, browser_hover, browser_drag, browser_wait_for, browser_get_page_text, browser_take_screenshot, browser_file_upload, browser_attach_image, browser_tabs, browser_console_messages, browser_network_requests, browser_computer, browser_handle_dialog, browser_gif_creator, browser_resize… Each MCP process is its own session (GS_BROWSER_SESSION, a random id by default): the extension gives it its own tab, so several agents can drive the same Chrome at once without stepping on each other. Every tool also takes an optional tabId: open more tabs with browser_tabs {action:"new", background:true} to run several tasks of the same agent in parallel. When the process exits, its tab is freed right away and closed after 2 minutes, so tabs don't pile up. browser_batch {actions:[{name, input}, …]} runs several tools in ONE call, in order (navigate, type, click, wait, read): one model round trip instead of many, and only the last step returns a snapshot. It stops at the first error, pending confirmation, captcha, dialog, or a stale ref after the page changed — after navigating, target with a selector or text. Native dialogs (confirm/alert/prompt/beforeunload) are dismissed by default and reported in the action's answer; call browser_handle_dialog {accept:true} before the action to accept the next one, or on one left open. No tool hangs on an open dialog.

Env: GS_BROWSER_TOKEN, GS_URL (default https://www.ghosty.studio), GHOSTY_BROWSER_VIA=native|gs, GS_BROWSER_CLIENT (name shown in the panel). Docs: https://www.ghosty.studio/docs/navegador