Skip to content

Remote MCP Server Setup for AI Clients

This guide explains how to point an MCP-capable AI client at a hosted Model Context Protocol server that wraps a scraping and data-extraction API. The server runs remotely over HTTPS, so there is no local service to install. You configure a client, supply an API key in the endpoint URL, and the assistant can then call scraping, search, e-commerce, YouTube, and account tools on demand.

What MCP does

The Model Context Protocol is an open standard for letting AI clients call external tools. A client that speaks MCP can connect to any compatible server without a bespoke integration. In this setup, the server acts as a tool provider that gives the assistant live web access for scraping, searching, and structured extraction. The client discovers the available tools during the MCP handshake and decides which ones to call based on the task.

How the hosted server works

The server exposes the underlying API as MCP tools. Connect a client such as Claude Desktop, Cursor, or ChatGPT and the assistant can scrape pages, take screenshots, run Google/Amazon/Walmart/YouTube searches, and send prompts to ChatGPT or Gemini, all billed to your account. The hosted endpoint is at {{MCP_BASE_URL}}/mcp. Your client connects directly over HTTPS; the API key is appended as a query parameter when you configure the client, as described in the authentication section.

Example workflows

Two patterns show how the tools combine. In a research workflow, you can ask the assistant to find the three best Python libraries for PDF parsing in 2026 and summarize the trade-offs. The assistant calls fast_search to find candidates, then get_page_text on each result to read the documentation and GitHub READMEs, without you pasting URLs. In a price-monitoring workflow, you can paste a list of ASINs and ask the assistant to check the current Amazon price for each and flag anything that dropped more than 10% this week. The assistant calls get_amazon_product_details per ASIN, compares against earlier numbers, and returns a summary. The assistant chooses tools from the descriptions it sees at handshake, so you can combine them however the task requires.

Before you connect

Three things need to be in place. First, an account with an API key, which you can retrieve from the key management page at {{APP_BASE_URL}}/account/manage/api_key. Second, a client that can act as an MCP host: Claude Desktop, Cursor, ChatGPT on Plus, Pro, Business, or Enterprise, or any other program that supports MCP over streamable HTTP. Third, if your client reaches the server through the mcp-remote wrapper, which is how Claude Desktop and Cursor do it, you need Node.js installed on the same machine.

Connecting your AI client

The server URL is the same for every client; only the configuration format differs. The subsections below cover Claude Desktop, Cursor, ChatGPT, and other MCP clients.

Claude Desktop

Open claude_desktop_config.json (Claude → Settings → Developer → Edit Config) and add the {{OWNER_BRAND_NAME}} entry under mcpServers. The exact JSON is provided in protected module code0001 and must be reproduced verbatim. Put your real key where the placeholder sits, then quit Claude Desktop entirely and relaunch it. The connector appears in the + menu next to the prompt input, under Connectors; enable it there and it remains usable in every conversation. If the model might otherwise pick its own browsing tools or another installed connector, name the server explicitly in the request, as in "use {{OWNER_BRAND_NAME}} mcp to find the top 5 Python PDF parsing libraries."

Cursor

In Cursor, go to settings, locate the MCP area, and choose + New MCP Server ("Add a Custom MCP Server"). The editor opens its mcp.json file; insert the entry from protected module code0002 with your own key in place of the placeholder. This relies on the same mcp-remote wrapper as Claude Desktop. Once the file is saved, the MCP panel does not refresh by itself: open a different settings section and come back to MCP, and {{OWNER_BRAND_NAME}} should be listed. Turn it on and a green status indicator shows the tools are ready for Cursor's agent. If the agent could otherwise reach built-in browsing, a different MCP server, or the {{OWNER_BRAND_NAME}} CLI skill when installed, say "use {{OWNER_BRAND_NAME}} mcp" in the request, for example "use {{OWNER_BRAND_NAME}} mcp to extract product prices from this Amazon URL." The default target is the global file at ~/.cursor/mcp.json, which applies to every project; to limit the connector to one repository, place the identical JSON in .cursor/mcp.json at that project's root.

ChatGPT

ChatGPT supports custom MCP servers as apps. MCP custom apps are available on Plus and Pro (read/fetch tools only) and on Business, Enterprise, and Edu (full MCP including write actions, in beta). To connect {{OWNER_BRAND_NAME}}, switch on developer mode via Settings → Apps → Advanced Settings → Developer mode; on a Business or Enterprise workspace an administrator may need to permit it beforehand through Workspace Settings → Permissions & Roles → Connected Data. Next, open Settings → Apps → Create to bring up the New App dialog. Give the app the name {{OWNER_BRAND_NAME}}, since that is the name you will type in prompts. In the Connection field, keep the toggle on Server URL and paste the MCP endpoint: {{MCP_BASE_URL}}/mcp?api_key=YOUR_API_KEY (replace YOUR_API_KEY with your real key). Leave Authentication set to No Auth; the API key is already embedded in the URL. Tick the I understand and want to continue checkbox, which acknowledges the security warning ChatGPT shows for any custom MCP server, then press Create. Creation takes a few seconds. A dialog titled Add {{OWNER_BRAND_NAME}} to ChatGPT appears with permission and data-use notes; press Connect to finish the setup. The app then appears under Settings → Apps → Enabled Apps with a Dev label, and on a Business or Enterprise workspace an administrator publishes it for the workspace. Once connected, invoke {{OWNER_BRAND_NAME}} in any chat by typing @{{OWNER_BRAND_NAME}} in your prompt, which routes the request straight to the MCP app, for example: "@{{OWNER_BRAND_NAME}} get the current price of ASIN B00P7VW2H6". Alternatively, click the + button next to the prompt input → More → select the app from the list. Custom MCP app setup has to be done from chatgpt.com on the web.

Other MCP clients

The endpoint works with any client that supports MCP over streamable HTTP. A stdio-only client can use the server through an adapter: place mcp-remote between the client and the remote URL. The Claude Desktop and Cursor configurations demonstrate that setup. For a direct integration, the Custom Python Client tab at {{MCP_BASE_URL}} provides an httpx example. Its sequence initializes MCP, requests the tool catalog and invokes a tool.

Available MCP tools

The server exposes the tools grouped below by capability. tools/list is the discovery step: it returns the parameter definitions for each tool. AI assistants commonly use those definitions to display inputs beside a tool invocation.

HTML API tools

The HTML API group covers page text, raw HTML, selector-based extraction, screenshots, and file downloads. The exact tool names and descriptions are in the HTML API tool table (table0001). premium_proxy, custom_google and country_code are common to all five HTML tools for proxy selection and geographic targeting. Use stealth_proxy only with the four browser-rendered tools. get_file downloads raw files without a browser, so it does not support the JavaScript-dependent stealth option. When to use premium_proxy and stealth_proxy: start without either. If the site blocks you (CAPTCHA page, empty result, or error response), retry with premium_proxy: true. If you are still blocked, escalate to stealth_proxy: true. The ladder goes cheap to expensive, so only step up when you need to. See credit costs for exact pricing per option.

Search and AI tools

The search and AI group covers general web search, specialized Google search types, and prompts to ChatGPT or Gemini. The exact tool names and descriptions are in the search and AI tool table (table0002).

E-commerce tools

The e-commerce group covers Amazon search, product details, pricing and seller offers, plus Walmart search and product details. The exact tool names and descriptions are in the e-commerce tool table (table0003).

YouTube tools

The YouTube group covers search, video metadata, and subtitles. The exact tool names and descriptions are in the YouTube tool table (table0004).

Account tool

The account group contains a single usage tool. The exact tool name and description are in the account tool table (table0005).

Authentication

Put the API key in the connection URL as the api_key query value: {{MCP_BASE_URL}}/mcp?api_key=YOUR_API_KEY. The server reuses that key when dispatching each API call. This authentication contract uses neither request-header authentication nor OAuth. A connection URL containing an API key is a secret, so keep it out of repositories. Client configurations in a home directory or workspace can expose the key to anyone with file access. Use environment variables for scripts and replace the key through the account dashboard after exposure.

Credits and limits

In the reference credit schedule, MCP is a transport layer rather than a second billable operation: get_page_html carries the same charge as an HTML API request. There is no separate MCP surcharge. Per-endpoint credit costs appear in the matching API reference (HTML API, Google API, Amazon API, and so on). Two concurrency ceilings apply: the account allowance and the hosted server capacity shared across customers. A busy shared service can return HTTP 429 Too Many Requests. Increase the retry delay exponentially when the client exposes that response; the account limit still applies.

Troubleshooting

Missing API key: tool calls return "{{OWNER_BRAND_NAME}} API key is missing in the remote mcp url." The ?api_key= query parameter is missing or empty in the URL your client is using. Edit the client config so the URL contains the full key (?api_key=YOUR_KEY) and reload the client. Invalid API key (HTTP 401): tool calls return "Error 401: Invalid api key." The key is present but rejected, invalid, revoked, or out of credits. Check the key at {{APP_BASE_URL}}/account/manage/api_key and your credit balance. Client cannot find the server or handshake never completes: for clients using the mcp-remote wrapper (Claude Desktop, Cursor), make sure Node.js is installed and npx is on your PATH. Run npx mcp-remote "{{MCP_BASE_URL}}/mcp?api_key=YOUR_KEY" in a terminal. A healthy run prints Connected to remote server using StreamableHTTPClientTransport followed by Proxy established successfully. If you do not see those lines, the issue is local (Node.js missing, outbound network blocked); if you do see them but your client still cannot connect, the issue is in the client's MCP config.

Resources

Useful references include the Model Context Protocol specification, the Scraping API dashboard for managing your API key, viewing usage, and upgrading your plan, and the Knowledge Base for searchable troubleshooting articles.

HTML API tool table
ToolWhat it does
get_page_textScrapes a URL and returns clean text or Markdown. Use premium_proxy or stealth_proxy options to bypass advanced blocking.
get_page_htmlRetrieves the full, raw HTML content of a webpage, essential for tasks that require parsing the DOM or analyzing site structure.
extract_page_dataExtracts specific data from a webpage using CSS or XPath selectors. Define extraction rules as a JSON object to pull exactly the data you need.
get_screenshotCaptures a visual screenshot of a webpage. Can capture the full page or a specific element identified by a CSS selector.
get_fileA versatile tool to fetch any file from a URL, including images, PDFs, or other documents.

All five accept premium_proxy, custom_google, and country_code. The first four also accept stealth_proxy; get_file does not, because stealth_proxy requires JavaScript rendering and get_file runs without it.

Search and AI tool table
ToolWhat it does
fast_searchThe primary tool for web searches. Returns clean organic results and top stories in under one second at a lower credit cost. Use this first for any general search.
get_google_search_resultsA fallback search tool for specialized search types: news, maps, Google Lens, shopping, image search, or Google AI mode.
ask_chatgptSends a prompt to a ChatGPT model, with an option to enhance the response with live web search results for up-to-date answers.
ask_geminiSends a prompt to Gemini and returns the generated answer, including citation objects when Gemini supplies them.

fast_search is the primary general search tool; get_google_search_results is the fallback for specialized search types.

E-commerce tool table
ToolWhat it does
get_amazon_search_resultsScrapes Amazon search result pages for a given query, providing a structured list of products to power e-commerce analysis.
get_amazon_product_detailsFetches clean, detailed information for any Amazon product using its ASIN, including pricing, ratings, and review counts.
get_amazon_pricingReturns the current price and the full list of seller offers for an ASIN — use it to see who holds the buy box and at what price.
get_walmart_search_resultsPerforms a product search on Walmart and returns a structured list of results with filtering and sorting capabilities.
get_walmart_product_detailsRetrieves structured data for a specific Walmart product, including price and availability, localized by store or zip code.

Amazon tools use ASIN; Walmart product details are localized by store or zip code.

YouTube tool table
ToolWhat it does
get_youtube_search_resultsSearches YouTube videos, channels, or playlists with extensive filters for duration, upload date, quality (HD/4K), and more.
get_youtube_video_metadataRetrieves comprehensive metadata for a specific video: title, description, view count, likes, keywords, and format details.
get_youtube_video_subtitlesRetrieves the subtitle track for a single video. Choose a language, and use subtitle_origin to pick between auto-generated and uploader-provided captions.

Subtitle selection requires a language and subtitle_origin choice.

Account tool table
ToolWhat it does
get_scrapingbee_usageReads real-time credit and concurrency usage for the authenticated account, so the assistant can check your remaining balance before starting a large job. Rate-limited to 6 calls per minute.

Rate-limited to 6 calls per minute.

Frequently asked questions

Do I need to install anything locally to use the remote MCP server?

No local server is required. The server is hosted at {{MCP_BASE_URL}}/mcp and your client connects over HTTPS. However, if your client uses the mcp-remote wrapper (Claude Desktop and Cursor do), Node.js must be installed locally so npx can run the wrapper.

How is the API key passed to the server?

The key is appended to the server URL as the api_key query parameter, for example {{MCP_BASE_URL}}/mcp?api_key=YOUR_API_KEY. There is no header-based or OAuth flow. The server forwards the key to the underlying API on each tool call.

What is the difference between a missing key and an invalid key error?

A missing key error means the ?api_key= parameter is absent or empty in the URL your client uses; fix the client config and reload. An invalid key error (HTTP 401) means the key is present but rejected, for example because it is invalid, revoked, or the account is out of credits; check the key and your balance.

Does using MCP cost more than calling the API directly?

No surcharge is added. Each MCP tool call consumes credits at the same rate as the underlying API endpoint it wraps. For example, calling get_page_html via MCP costs the same as calling the HTML API directly.

Why does get_file not accept stealth_proxy?

stealth_proxy requires JavaScript rendering, and get_file runs without a browser because raw file downloads do not need one. The other HTML API tools accept stealth_proxy.

How do I make sure my assistant uses the MCP server instead of built-in web tools?

Include the phrase "use Scraping API mcp" in your prompt. For ChatGPT, type @Scraping API in the prompt to route the request directly to the MCP app.

{
  "mcpServers": {
    "scrapingbee": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "{{MCP_BASE_URL}}/mcp?api_key=YOUR_API_KEY"
      ]
    }
  }
}
{
  "mcpServers": {
    "scrapingbee": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "{{MCP_BASE_URL}}/mcp?api_key=YOUR_API_KEY"
      ]
    }
  }
}