@unblockingapi/mcp

Official Model Context Protocol server for UnblockingAPI. Gives AI agents (Claude, Cursor, OpenClaw, …) the ability to fetch bot-protected and JavaScript-heavy web pages and run structured Google searches — all through UnblockingAPI's anti-detection engine and rotating residential proxies.

Tools

More tools (e.g. Ahrefs website authority) will be added as additional API templates go live.

A note on rendering & block_assets

When you set render: true, the page is rendered in a real browser and JavaScript always executes — so SPAs and dynamic content come back fully rendered. By default the renderer skips downloading CSS, images, fonts, and media (block_assets defaults to true on renders) for speed and lower cost; this does not affect the DOM/text you get back. Pass block_assets: false only when you actually need those assets (e.g. image URLs or a visually complete page).

Related MCP server: pure.md MCP Server

Setup

You need an UnblockingAPI key — get one at https://unblockingapi.com.

The server runs over stdio (launched with npx). Desktop/CLI clients that spawn a local process (Claude Desktop, Claude Code, Cursor, …) use it directly. Web clients that only accept a remote URL (ChatGPT) need it bridged to HTTP — see the ChatGPT section.

Install by prompt

Agentic clients that can run commands or edit their own config (Claude Code, Cursor's agent, Cline, etc.) can install the server when you just ask. Paste a prompt like:

"Install the @unblockingapi/mcp MCP server. My API key is sk_xxx. It runs over stdio via npx -y @unblockingapi/mcp and needs the env var UNBLOCKINGAPI_KEY."

In Claude Code that's enough — it will run the right claude mcp add for you. In Cursor / Cline the agent will add the entry to your mcp.json. After it finishes, start a new chat (or reload MCP servers) so the tools load.

GUI apps without an agent that can edit config — Claude Desktop and ChatGPT — can't self-install from a prompt; use the manual steps below.

Claude Desktop

Edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\), then restart Claude:

{
  "mcpServers": {
    "unblockingapi": {
      "command": "npx",
      "args": ["-y", "@unblockingapi/mcp"],
      "env": { "UNBLOCKINGAPI_KEY": "your_api_key_here" }
    }
  }
}

Claude Code (CLI)

claude mcp add unblockingapi \
  -e UNBLOCKINGAPI_KEY=your_api_key_here \
  -- npx -y @unblockingapi/mcp

Cursor

Add to ~/.cursor/mcp.json (or a project .cursor/mcp.json) — same shape as the Claude Desktop block above.

ChatGPT

ChatGPT (Developer mode → Connectors) only accepts remote MCP servers reachable over a public HTTPS URL — it can't spawn a local npx process. Bridge this stdio server to HTTP with supergateway:

UNBLOCKINGAPI_KEY=your_api_key_here \
  npx -y supergateway --stdio "npx -y @unblockingapi/mcp" --port 8000
# exposes an MCP endpoint at http://localhost:8000/sse

Then expose it publicly (ChatGPT can't reach localhost) with a tunnel, e.g. cloudflared tunnel --url http://localhost:8000 or ngrok http 8000, and add the resulting https://…./sse URL under Settings → Connectors in ChatGPT.

For production ChatGPT use you'll want a hosted HTTPS endpoint rather than a local tunnel. A first-class hosted/streamable-HTTP transport is on the roadmap.

Environment variables

Usage examples

Once connected, ask your agent things like:

• "Fetch the rendered HTML of https://example.com using a German proxy." → unblock_fetch(url, render: true, location: "de")

• "Get the Google results for 'best running shoes' in the US." → google_search(q: "best running shoes", location: "us")

• "Pull the details of this idealista listing: https://www.idealista.com/inmueble/111072490/" → idealista_property(url: "https://www.idealista.com/inmueble/111072490/")

unblock_fetch parameters

• url (required) — HTTP/HTTPS URL. Media/binary files are rejected.

• render — render with a headless browser (runs JS). Default false.

• location — 2-letter country code for the proxy (us, gb, de, …).

• wait — render-only. Comma-separated wait steps (max 5): a leading load event (domcontentloaded|load|networkidle), then CSS selectors or networkidle:<ms> / domstable:<ms> strategies. e.g. domcontentloaded,h3.

• block_assets — render-only. Defaults to true on renders (skips CSS/images/fonts/media for speed; JS still runs). Pass false to also download those assets. See note above.

• remove_scripts / remove_stylesheets / remove_svgs — strip those tags from the returned HTML.

Development

npm install
npm run dev        # tsx watch
npm run typecheck  # tsc --noEmit
npm run build      # compile to dist/

License

MIT