How to Connect AI Agents to Real-Time Amazon Data: The Complete Runnable Guide

An AI agent is, at minimum, a language model wrapped in a loop: it observes, reasons, acts, and observes again. What distinguishes a data-connected agent from a bare model is a single property — grounding. A grounded agent can resolve a claim against an external source at the moment of use; a bare model can only resolve it against its parameters, which were fixed at training time. On stable knowledge the two are indistinguishable. On commercial questions — prices that change hourly, reviews written yesterday, rankings that shift with every campaign — the bare model answers with the confidence of knowledge and the substance of recollection, and the reader has no instrument for telling the two apart.

This post gives you the complete, runnable connection: the MCP endpoint, the three-step first test with exact tool-call payloads, Claude Code / Claude Desktop registration, and a Python reference implementation with the verified expected outputs.

Why: the experiment that measures the gap

On 2026-10-12 we ran the same prompt — "What are the main complaint themes in AirPods Pro 3 reviews?" — against the data layer in two conditions.

In the grounded condition, search_amazon found the product (ASIN B0FQFB8FMG, Apple AirPods Pro 3, 4.4 stars across 16,350 ratings), and get_amazon_reviews with filterByStar=critical returned 1,392 critical reviews out of 4,268 total, the newest dated October 10, 2026 — two days before the test. Four dominant themes, each traceable to a verified purchase: software regressions after updates, noise cancellation underperforming expectations, units failing within 4–6 months, incoming-call handling problems. Every review carried a verified-purchase flag and a helpful-vote count.

In the ungrounded condition — the same prompt answered from parametric knowledge alone — the answer named plausible themes (noise cancellation, fit, price) but supplied no counts, no dates, no verification markers, and carried a contamination risk from the previous generation: memories of the older AirPods Pro 2 leaking into an answer about the Pro 3, with nothing in the output letting the reader detect the substitution.

The argument follows: for commercial questions, the deficit of an ungrounded agent is not eloquence but verifiability. A claim without a number, a timestamp, or a source is not a finding; it is a hypothesis wearing a finding's clothes.

The three integration patterns

The selection rule is mechanical: if the agent talks to a human, prefer MCP; if a machine runs on a schedule, prefer REST; if the agent lives in an IDE, prefer Skill. The pricing is uniform underneath: product data at 1 credit per page, review data at 5 credits per page; MCP tool calls cost the same as their REST equivalents — the protocol changes who initiates the call, not what the data costs.

The endpoint is one public HTTPS server: https://mcp.pangolinfo.com/mcp — 20 tools (19 data + 1 free self-inspection tool), authenticated with Bearer <API key>.

Step 1 — Get your free key

Signup at tool.pangolinfo.com includes a free tier of 60 credits per month, recurring (reset monthly). The entire three-step test below costs single-digit points — run it end to end before any payment decision.

Step 2 — Register the MCP server in your client

Claude Code — one-liner:

claude mcp add --transport http pangolinfo https://mcp.pangolinfo.com/mcp --header "Authorization: Bearer YOUR_KEY"

Claude Desktop — config file snippet (keep the key out of any text you share):

{
  "mcpServers": {
    "pangolinfo": {
      "url": "https://mcp.pangolinfo.com/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_KEY"
      }
    }
  }
}

Meta Muse takes the same endpoint as a custom connector (Settings → Connectors → Add custom connector, key through the secure credential flow) — see the Meta Muse installment for the full walkthrough.

Step 3 — Run the three-step test

Discovery. The MCP tool is search_amazon:

{
  "keyword": "wireless earbuds",
  "site": "amz_us",
  "page": 1
}

Expected (verified 2026-10-12): five live ASINs, including B0FQFB8FMG — Apple AirPods Pro 3, 4.4 stars, 16,350 ratings.

Complaint mining. The MCP tool is get_amazon_reviews, with the review endpoint as the REST equivalent (same filters, 5 credits per page):

{
  "asin": "B0FQFB8FMG",
  "site": "amz_us",
  "pageCount": 1,
  "filterByStar": "critical",
  "sortBy": "helpful"
}

Expected (verified 2026-10-12): 1,392 critical reviews out of 4,268, newest dated 2026-10-10, four traceable themes — software regressions after updates, weak noise cancellation, 4–6 month hardware failures, incoming-call handling problems.

Python reference implementation

Drives the same three steps over the standard MCP transport. Install:

pip install mcp httpx
export PANGOLIN_API_KEY="your-key"

import asyncio, json, os
from mcp import ClientSession
from mcp.client.streamablehttp import streamablehttp_client

ENDPOINT = "https://mcp.pangolinfo.com/mcp"

def show(label, result):
    print(f"\n=== {label} ===")
    for block in result.content:
        text = getattr(block, "text", "")
        print(text[:1200])   # first screen of the payload

async def main():
    headers = {"Authorization": f"Bearer {os.environ['PANGOLIN_API_KEY']}"}
    async with streamablehttp_client(ENDPOINT, headers=headers) as (read, write, _):
        async with ClientSession(read, write) as session:
            await session.initialize()

            # Step 1 — discovery: five live ASINs, target B0FQFB8FMG
            disc = await session.call_tool("search_amazon", {
                "keyword": "wireless earbuds", "site": "amz_us", "page": 1})
            show("discovery", disc)

            # Step 2 — complaint mining: 1,392 critical of 4,268, newest 2026-10-10
            crit = await session.call_tool("get_amazon_reviews", {
                "asin": "B0FQFB8FMG", "site": "amz_us",
                "pageCount": 1, "filterByStar": "critical", "sortBy": "helpful"})
            show("critical reviews", crit)

            # Step 3 — sanity: do the numbers match the expected verification?
            # Expect: 5 ASINs incl. B0FQFB8FMG; 1,392/4,268 critical; newest 2026-10-10.

asyncio.run(main())

Honest scope note. The tool surface and payloads above were verified live on 2026-10-12 against this exact endpoint; the Meta Muse connector path was verified end to end by our team. The Python path above runs over the standard MCP Streamable HTTP transport, which any spec-compliant client library implements. The REST pattern is: host api.pangolinfo.com, Authorization: Bearer <key>, same per-page credit rates (1 for product data, 5 for reviews) — endpoint reference is in the docs, not reprinted here.

Platform survey: where each agent meets the data

• Meta Muse (launched 2026-09-08, US/Canada) — custom MCP connector; verified end to end. Walkthrough: Meta Muse guide.
• Claude — custom connectors / Claude Code MCP config. Auth-surface breakdown: Claude Code guide.
• ChatGPT — Custom GPT Actions, developer-mode MCP support. Installment forthcoming.
• Tencent Workbuddy (launched 2026-03-09, 8.85M MAU first month, domestic + international editions) — open platform + skill ecosystem. Installment forthcoming.
• IDE agents (Cursor-class) — MCP servers or skill packages inside the editor. Installment forthcoming.
• n8n / Make — scheduled REST workflows for continuous monitoring. Installment forthcoming.

For platforms whose connection flow our team has not yet verified end to end, the guide states the access surface honestly and defers the click-by-click procedure to the corresponding installment.

Cost

The Amazon Scraper API: product data 1 credit/page, review data 5 credits/page. MCP tool calls bill at the same rates. Free tier: 60 credits/month, recurring — the three-step test costs single-digit points.

Security: key management

1. Never paste the key into conversation text that may be logged or shared — use the client's credential field, an MCP Authorization header, or a vault.
2. Issue one key per environment or teammate so usage is attributable.
3. Treat read-only research as read-only: hold connectors to read-only or approval-on-write policies where the client offers them.

FAQ

Is the free key enough? Yes — 60 credits/month; the full test costs single-digit points.

MCP, REST API, or Skill? Follow the agent: conversational assistants → MCP; scheduled automation and batch jobs → REST API; coding agents (Claude Code, OpenClaw) → Skill. Same data, same pricing underneath.

Region restrictions? The agents have them, the data does not. Meta Muse is US/Canada only today; Workbuddy ships separate domestic and international editions.

Key safety in a connector flow? Credential field or header — never chat text. Rotate on schedule.

One key for every agent? Yes. Separate keys per agent are recommended for attribution, not required for function.

Series navigation

This guide is the hub: 1 pillar + 6 sub-guides (Meta Muse, Claude Code, ChatGPT, Tencent Workbuddy, IDE agents, n8n/Make). Each installment reuses the same three-step first test from above, so results are comparable across agents. Start with the free key, run the test, then pick the agent that fits your workflow.

Originally published on the Pangolinfo blog.