SuperAPI AI Directory

mcp

🎖️ 📇 ☁️ 🏠 - Web scraping that escalates from direct fetch to rotating proxy to full browser, solving anti-bot challenges and reporting which step worked and its cost.

Open website

About this tool

FourA MCP is an MCP server for the FourA web scraping API, providing AI agents with reliable web access capabilities. The user gives the agent a public URL, and FourA returns the content, whether the page requires a fast HTTP request, proxy rotation, or a full browser session.

Purpose This server is used to let AI agents scrape public web content, covering scenarios such as product information extraction, article cleaning, price monitoring, endpoint checking, and batch URL scraping. It connects to Claude Desktop, Claude Code, Cursor, Windsurf, VS Code, or any MCP client through the MCP protocol.

Main capabilities It provides four tools: foura_auto is the default smart scrape, returning content once a URL is entered, automatically selecting the scraping method and handling common blocks; foura_single performs a single fast HTTP request, suitable for static pages, JSON APIs, and server-rendered HTML; foura_proxy sends requests through rotating proxies and automatically retries on failure, suitable for targets where direct connections are blocked or a specific exit country is required; foura_browser launches a real browser session, executes JavaScript, completes DOM rendering, and returns cookies, suitable for single-page applications, lazy-loaded content, or anti-scraping challenges that require a real browser to pass. All tools are marked with readOnlyHint: true and openWorldHint: true. Each response contains human-readable text content and a structured JSON object structuredContent that conforms to the output schema. The result reports credits (the credits consumed by this call, also billed on failure), request_id (the call identifier), and exitClass (premium or standard). Six built-in prompt workflows are displayed under /prompts in the MCP client, orchestrating one or more tools without requiring the user to explain step by step.

Usage A FourA API key is required, in the format pk_live_..., created or viewed at foura.ai/dashboard/#api-keys. Local stdio mode is recommended for Claude Desktop. Configuration example: claude mcp add foura -- npx -y @fouradata/mcp (FOURA_API_KEY must first be set in the environment variables). The npx command downloads the package on first startup and runs it as a subprocess of the MCP client, with no global installation required. The same JSON configuration applies to different MCP clients; only the configuration file location differs. Claude Desktop must be fully quit (Cmd+Q on macOS) before editing the configuration file, otherwise the app will overwrite the changes with the in-memory configuration when it exits. Hosted mode uses a Streamable HTTP endpoint, and the client must support that transport method; Claude Desktop can use the stdio configuration above, or bridge the hosted endpoint through mcp-remote. stdio mode uses FOURA_API_KEY, and the hosted endpoint uses the Authorization: Bearer pk_live_... request header. One key authenticates all four tools.

Applicable scenarios and limitations It is suitable for AI agent workflows that need to obtain content from public web pages. Private targets are rejected; requests to private or reserved addresses are blocked at the MCP layer, and only public internet hosts are forwarded. Rate limits are enforced per service by the FourA API, involving concurrency and requests per minute. When using it, you should only access public content that you are authorized to access, and comply with the target website's terms and applicable laws.

Other details foura_single and foura_proxy support selecting a browser profile (browser, os, version, profile). The full catalog is publicly available at api.foura.ai/api/profiles, with no key required. foura_proxy returns the encoded ID of a successful proxy, which can be passed to foura_single.proxy or foura_browser.proxy to reuse the same exit. By default, the full response body is returned inline; when offload_large: true is passed, responses of 50 KB and above are returned as a resource_link, the resource is retained for one hour, and only the API key that created the resource can read it back. Error responses carry a structured envelope containing fields such as code and message, making it easy for agents to implement retry logic.