Amazon Products MCP Server

An MCP server for searching and getting product details from Amazon. Supports 22 regional Amazon sites, extracts prices, ratings, features, availability, and product variants. Uses a headless Chrome browser to render pages and handle Amazon's anti-bot protections. Communicates over stdio and works with any MCP-compatible client such as Claude Desktop or Claude Code.

Getting Started

Requirements

• Google Chrome or Chromium installed (the server launches it in headless mode)

  On Linux, avoid snap-packaged Chromium (Ubuntu's default chromium package). It is known to hang indefinitely — both on graceful CDP tab-close and on OS-level process reaping after a forced kill — when launched as a descendant of certain parent processes (observed running under Claude Desktop's Electron process tree). The hang was 100% reproducible under that setup and never reproduced with the same code against a non-snap Chrome/Chromium build. Install Google Chrome or a non-snap Chromium build instead, and point CHROME_EXEC_PATH at it if it isn't picked up automatically.

Configuration

Install from source

Requires Go 1.24+.

go install github.com/jbeshir/mcp-servers/amazon-products/cmd/amazon-products-mcp@latest

Docker

The Docker image bundles a headless Chrome shell, so no local Chrome installation is needed.

docker build -t amazon-products-mcp ./amazon-products
docker run -e AMAZON_REGION=us amazon-products-mcp

Claude Desktop

Add to your Claude Desktop configuration (claude_desktop_config.json):

{
  "mcpServers": {
    "amazon-products": {
      "command": "/path/to/amazon-products-mcp",
      "env": { "AMAZON_REGION": "us" }
    }
  }
}

Claude Code

claude mcp add amazon-products -- env AMAZON_REGION=us /path/to/amazon-products-mcp

Tools

Supported Regions

Key Concepts

• ASIN -- Amazon Standard Identification Number. A 10-character alphanumeric ID (e.g. B09CDRVQZC) that uniquely identifies a product on Amazon. Used as the primary key for get_product_details.
• Regions -- Each Amazon region maps to a specific domain (e.g. us maps to amazon.com, de maps to amazon.de). The region determines which site is scraped and which currency prices are reported in. A default region is set via AMAZON_REGION and can be overridden per request.
• Variants -- Product variations along dimensions like size, color, or style. Each variant dimension lists its options with states: SELECTED (currently viewed), AVAILABLE (can be selected), or UNAVAILABLE (out of stock). Each available option includes an ASIN for fetching that specific variant.
• WAF challenges -- Amazon's Web Application Firewall sometimes returns a 503 page with a JavaScript challenge instead of the real content. The browser handles this by navigating to the Amazon homepage first to acquire cookies, then retrying the target URL with a 3-second delay between attempts (up to 3 attempts).
• CAPTCHA detection -- If Amazon presents a CAPTCHA page, the server detects it and returns an error immediately rather than timing out. This is a distinct condition from WAF challenges and cannot be bypassed automatically.
• Anti-detection -- The headless browser overrides JavaScript properties (navigator.webdriver, navigator.plugins, window.chrome) and uses a realistic user agent and window size to reduce the chance of being identified as automation.

Architecture Overview

graph LR
    Client["MCP Client<br/>(Claude Desktop, Claude Code)"]
    Server["amazon-products-mcp<br/>MCP Server (3 tools)"]
    Scraper["Datasource<br/>region-specific"]
    Browser["Headless Chrome<br/>(chromedp)"]
    Amazon["Amazon<br/>22 regional sites"]

    Client -- "JSON-RPC over stdio" --> Server
    Server --> Scraper
    Scraper --> Browser
    Browser -- "HTTPS" --> Amazon

The server has three internal layers:

• cmd/amazon-products-mcp -- Entry point. Reads AMAZON_REGION from the environment, creates the shared browser instance and MCP server, and starts the stdio transport.
• internal/server -- Registers 3 MCP tools, resolves the target region per request, and formats responses as JSON.
• internal/scraper -- Two components: Datasource builds Amazon URLs for a given region and parses returned HTML into structured data. Browser manages a shared headless Chrome instance, handles WAF challenges/retries, and waits for page rendering to complete.

Data Flow

sequenceDiagram
    participant C as MCP Client
    participant S as MCP Server
    participant B as Headless Browser

    C->>S: search_products(query, region)
    S->>B: Fetch(searchURL)

    Note over B: Open new tab
    B->>B: Navigate to Amazon homepage
    Note over B: Acquire WAF cookies
    B->>B: Navigate to search URL
    B->>B: Wait for DOM stability

    alt WAF 503 challenge
        B->>B: Wait 3s, retry (up to 3x)
    end
    alt CAPTCHA detected
        B-->>S: ErrCAPTCHA
        S-->>C: Error: CAPTCHA page
    end

    B-->>S: Rendered HTML
    S->>S: Parse HTML into products
    S-->>C: Product results (JSON)

Each request runs in its own browser tab for concurrency safety. The browser waits for either a specific CSS selector (#productTitle for product pages) or DOM stability (body content unchanged for 2 consecutive polls at 500ms intervals) before capturing HTML.