Amazon MCP

An MCP server for amazon.in — search, filter, read product pages, reviews, and pincode-based delivery info. No API key, no login, no paid scraper service.

Works with: any MCP client (Claude Desktop, Cursor, Hermes, etc.) · Marketplace: amazon.in · Python: 3.11+

Tools

Related MCP server: Amazon MCP Server

Quickstart

git clone https://github.com/thenabbu/amazon-mcp.git
cd amazon-mcp
python3 -m venv .venv
.venv/bin/pip install "mcp>=2" curl_cffi beautifulsoup4 lxml
.venv/bin/python amazon_mcp/server.py   # starts on stdio

Self-check without network:

.venv/bin/python amazon_mcp/server.py --demo

{"search": {"total": 65756, "items": 22, "ok": true}, "product": {"title": "HP Victus 15 Smartchoice...", "price": "₹1,08,990.00", "variants": 8, "dogpage": "DOGPAGE", "ok": true}, "reviews": {"dp1": {"n": 13, "avg": 4.0, "total": 95266}, "dp2": {"n": 8, "avg": 3.4, "total": 42}, "ok": true}, "pincode": {"validation": {"110001": true, "400001": true, "12345": false, "abcdef": false, "000000": false}}}

MCP client config

Claude Desktop / Cursor (claude_desktop_config.json or .cursor/mcp.json):

{
  "mcpServers": {
    "amazon": {
      "command": "/path/to/amazon-mcp/.venv/bin/python",
      "args": ["/path/to/amazon-mcp/amazon_mcp/server.py"]
    }
  }
}

Hermes:

hermes mcp add amazon --command /path/to/.venv/bin/python --args /path/to/amazon-mcp/amazon_mcp/server.py --connect-timeout 90

Example

get_product on a phone URL:

title:    Redmi 15C 5G Prime Edition Dusk Purple 6GB + 128GB | Massive 6000mAh Battery | ...
price:    ₹17,999.00   MRP: ₹31,999   savings: -44%
brand:    Redmi        seller: Dawntech Etel!
rating:   4.0 (744)    availability: In stock
images:   11 hi-res    variants: 4 (2 colors x 2 storages)
emi:      EMI starts at ₹633 per month. No Cost EMI available
offers:   Cashback up to ₹539 (Amazon Pay ICICI) / No Cost EMI / Partner Offers
exchange: With Exchange up to 15,500 off
details:  40 spec keys   breadcrumbs: Electronics > Mobiles > Smartphones

get_reviews on the same product:

avg 4.0, 744 ratings, histogram 5*54% 4*21% 3*8% 2*3% 1*14%
8 top reviews, e.g. "Battery lasts me 2 days" — Aditi, 16 October 2026 (verified purchase)

How it survives Amazon's bot wall

Amazon sits behind Akamai Bot Manager. Plain HTTP clients get a 503 or a proof-of-work challenge page. The server uses curl_cffi (real browser TLS fingerprint) and solves both known challenges automatically:

• the bm-verify interstitial (compute the small proof-of-work, POST /_sec/verify, retry)

• the transient "Continue shopping" captcha form (single GET, retry)

Plus polite pacing (1.5–3 s between requests) so the server stays under the radar.

Limits (honest)

• Deep review pagination needs login. Amazon redirects /product-reviews/ to sign-in for anonymous visitors (re-verified Oct 2026). Without cookies you get the ~13 top reviews from the product page + the full ratings histogram. get_all_reviews accepts the params and returns parsed reviews if you drop authenticated cookies into the session.

• Search results cap at ~100 pages (Amazon's own ceiling; deeper pages return empty).

• Prime badge / Climate Pledge are not rendered on amazon.in search pages — nothing to extract.

• A+ content is sometimes an empty header when Amazon lazy-loads the rest.

• Rate-limit yourself: heavy bursts will trigger challenges (the server auto-solves, but don't hammer).

Development

.venv/bin/python amazon_mcp/server.py --demo     # offline parser self-check
python test_hardening.py                          # input-validation asserts

Parser modules (search_parse, product_extract, reviews_fetch, pincode_common) are pure functions over saved HTML — fixtures under amazon_mcp/, so the demo runs offline.

License

MIT — see LICENSE.