MCP server
cdpfleet-mcp gives an AI assistant a real browser on the fleet through the Model Context Protocol: it picks the browser the site should see, browses through a residential IP (by country) or your own proxy, reads pages as accessibility snapshots, and can hand the browser to a person — to solve a captcha or type a code — through live view. Every session is recorded, so you can watch afterwards what the agent did.
Hosted: nothing to install
Step-by-step for Claude, ChatGPT, Claude Code, Codex and Cursor: MCP setup.
The same tools run at https://cdpfleet.com/mcp (MCP over Streamable HTTP). Your browsers are billed and limited by your plan exactly as from code.
- Claude, ChatGPT, Cursor, Claude Code…: add
https://cdpfleet.com/mcp; your assistant opens a cdpfleet window — sign in and click Allow (OAuth). Disconnect any time in the dashboard → API keys → Connected AI apps. - Without OAuth (scripts, headless agents): send
Authorization: Bearer <your API key>.
Each MCP connection gets its own browser; one left idle for 30 minutes is closed, which ends its browser session.
Or run it locally
Claude Code:
claude mcp add cdpfleet -e CDPFLEET_API_KEY=your-key -- npx -y cdpfleet-mcp
Claude Desktop, Cursor, Windsurf and other clients — add to the client's MCP configuration (claude_desktop_config.json, .cursor/mcp.json, …):
{
"mcpServers": {
"cdpfleet": {
"command": "npx",
"args": [
"-y",
"cdpfleet-mcp"
],
"env": {
"CDPFLEET_API_KEY": "your key"
}
}
}
}
Optional: CDPFLEET_PROXY sets the default proxy (your own proxy URL); without it the agent uses the cdpfleet residential proxy, billed per GB from your balance by tier ($5 free credit).
Tools
| Tool | What it does |
|---|---|
launch_browser | Starts a browser: browser (chrome, edge, brave, opera, whale, yandex, chromium, patchright, camoufox, firefox, webkit), headless (default headful), proxy, country (residential exit), record, and any other launch option. |
navigate | Opens a URL; returns the status, final URL and title. |
snapshot | The page (or one element) as an accessibility snapshot — roles, names, text and links, a fraction of the HTML's size (case). |
click, type, press, wait_for | Act on the page by CSS selector or by role and name from the snapshot. |
screenshot | A JPEG of the viewport or the whole page. |
evaluate | Runs JavaScript in the page and returns the result (data extraction). |
live_view | Hands the browser to a person: they watch or take control in the dashboard's Sessions page while the agent waits. |
close_browser | Ends the session; its recording appears under Sessions shortly after. |
list_browsers, account | The browsers and versions on offer; your usage, live sessions and residential balance. |
Good to know
- One browser at a time per MCP server; launching again closes the previous one.
- Billing is the same as for your code: thread time while the browser runs (a headful browser takes 2 threads), residential traffic per GB.
- Not built in: captcha solving — use
live_viewto let a person do it. - Built on the Node.js SDK; sources at github.com/cdpfleet/cdpfleet/sdk/mcp.