CloudBrowser

What CloudBrowser MCP server does

CloudBrowser exposes cloud-hosted Chrome through MCP so an AI client can operate a browser on the user’s behalf. Agents can open browsers, connect to them, visit requested URLs, inspect page content and elements, click controls, enter text, run scripts, and capture screenshots. Browser instances belong to the connected CloudBrowser account and follow its usage plan.

The service also supports saved browser sessions. A saved session can be listed, removed, or resumed later, restoring cookies and storage. Resuming does not restore open tabs. Remote desktop access lets a user watch a browser or take control while it is running.

How it works

The hosted MCP endpoint is https://mcp.cloudbrowser.ai and uses Streamable HTTP. A compatible client authenticates through OAuth 2.1 with PKCE and dynamic client registration. The user signs in to CloudBrowser and approves access; an API token can also be supplied in the client’s secret settings using the Authorization header.

A typical workflow is to inspect the workspace or active browsers, open or select a browser, connect to it, and then perform page operations. Disconnecting ends the control connection while leaving the browser running; closing the browser stops it. Resuming a saved session reuses an active browser when available, or opens a metered browser with a 180-second idle timeout.

Setup and configuration

Add https://mcp.cloudbrowser.ai as a custom MCP connector or HTTP server in the client. The documented setup paths cover Claude Desktop and claude.ai, Claude Code, Codex CLI, Cursor, VS Code, and Gemini CLI. ChatGPT can use the endpoint as a custom connector in developer mode. OAuth sign-in is required, and a CloudBrowser account is needed.

No API key needs to be copied during the normal OAuth flow. If using a CloudBrowser API token instead, keep it in the client’s protected secret configuration rather than placing it in prompts. The account’s CloudBrowser plan governs browser usage.

Tools and capabilities

• Manage browser instances with open_browser, close_browser, get_browsers, and get_connected_browsers.
• Connect to or disconnect from a browser without confusing control-session state with browser lifetime.
• Navigate pages, retrieve HTML content, inspect matching elements, click, type, evaluate scripts, and take screenshots.
• List, resume, and remove saved sessions.
• Start and stop remote desktop access.
• Display workspace information, including an optional connected-browser screenshot.
• Detect CAPTCHA pages and expose them for user handling.

Limitations and notes

Use browser actions only with websites and accounts the user owns or is authorized to use, and follow those sites’ terms. Page-content retrieval is intended for answering questions about the current page, not crawling or bulk extraction from third-party sites.

CloudBrowser browsers open without stealth fingerprints. Login prompts, MFA, verification walls, and CAPTCHAs remain with the user; the navigation flow stops rather than solving them. A resumed session restores cookies and storage but not tabs. Close metered browsers when finished, and do not retry an ambiguous browser allocation by changing its request ID.