Pricing

from $3.00 / 1,000 read operations

Amazon MCP Server - Product Data for AI Agents

Amazon MCP server for Claude, ChatGPT and Cursor. Search products and read live prices, buy-box offers, public reviews and bestseller rankings across US, UK, French and German stores, with no Amazon API key. Not affiliated with Amazon. By MrBridge (mr-bridge.com)

Pricing

from $3.00 / 1,000 read operations

Rating

0.0

(0)

Developer

MrBridge

Actor stats

Bookmarked

Total users

Monthly active users

9 days ago

Last modified

Changelog

All notable changes to this Actor. Dates are the build date. Versions are Apify build numbers, so they advance on every deploy and skip the ones that failed to build.

Unreleased

Added

Batch mode. A conventional run now performs one read and finishes, instead of starting a server nobody calls and timing out. The Input form takes an operation, a marketplace and its one conditional field; the run writes a single business item to the dataset and exits SUCCEEDED. This is what makes the Actor usable as a saved Task or on a schedule.
A task-results dataset view showing what a run fetched, alongside the existing receipts view used by Standby.

Changed

The Input schema declares real fields. It previously advertised an empty form and said the Actor took no run input, which is no longer true.
The dataset item written for a paid read now depends on the run mode: Standby keeps its technical receipt, a conventional run writes its result. A read bills by landing an item in the default dataset, so writing both would bill one call twice.
Store title, description and SEO copy left .actor/actor.json. apify push never wrote them to the live Actor, so the manifest was a second source of truth that only ever disagreed with the first.

Notes

Only the two read operations are offered in batch. The three action tools bill through tool-action and would additionally fire the synthetic dataset event.

Fixed

The batch run record no longer carries a free-form message. One failure message is built around the caller's own category slug, so logging the outcome whole would have persisted that slug beyond the run.
The output schema exposes one neutral dataset link instead of stating that the Actor produces no batch result. Two view-pinned outputs were tried first and were wrong: a view transforms every item rather than filtering, so each one rendered the other run's items as partial rows. The Console view selector does the job.
The results view names the product fields at the row root, which is where unwind puts them, rather than under a products. prefix.
The results view unwinds the products array, so a Task page shows one readable row per product. Display only: the run still writes one item and bills one read.
Batch input is closed to unknown fields, in the parser and in the published schema. An argument nobody wired up is refused rather than dropped and billed.

Unreleased (input form)

Changed

The Input form shows only Operation and Marketplace. The Actor is an MCP server first, and a form opening on a search box implied a visitor had to configure a query before they could connect a client. The MCP endpoint now leads the page.
query, category, page and maxResults are hidden rather than removed. They remain in the schema, in the parser and in every saved Example, and an API caller can still send them: only the form stopped showing them.
Operation and Marketplace values carry human labels: Amazon Search, Amazon Best Sellers, and each storefront alongside its domain.

Unreleased (taxonomy)

Fixed

A best sellers category that does not resolve now returns NOT_FOUND instead of PARSE_ERROR. The extractor always told the two apart, grid rendered with no cards being an unknown category rather than a layout change, but the fetch ladder converted every extractor exception into PARSE_ERROR and that judgement never reached the caller. Two costs: a caller was told their well-formed slug had hit a layout change, and the alarm that means "go read the extractor" fired on every typo. Measured on /gp/bestsellers/computers and /gp/bestsellers/toys, which Amazon answers with 200, no redirect, the page frame rendered, and the word undefined where the department name belongs.
On the vendor tier, that same verdict now stops the chain instead of failing over. A category that does not exist will not exist for the next vendor, and the second call buys the same answer at the same price.
The propagated NOT_FOUND carries the ladder's facts: tier, outbound attempts, ladder duration, queue and token waits, and the vendors actually tried. Passing the extractor's error through untouched had dropped all of them, so tool_call could no longer show that nothing escalated.

Only NOT_FOUND is propagated. Every other extractor exception, taxonomy or native, still reports PARSE_ERROR. The enrichment builds a new error rather than mutating the original, and inherits none of its details: the best sellers extractor throws with the caller's own slug in there, and telemetry outlives the run.

0.1.68 (2026-08-12)

Changed

Clarified the response returned when GET /mcp is used. The server supports Streamable HTTP through POST /mcp, but does not expose a standalone GET event stream.

0.1.67 (2026-08-12)

Fixed

A managed vendor that loses its own network budget is now cancelled, not merely outrun. The budget was a Promise.race against a timer, which bounded the wait and left the request on the wire until the ladder's global abort fired up to twenty seconds later, still billed by the vendor and still holding its rate-limit slot. Each attempt now carries its own AbortController, combined with the ladder's, and the socket dies at its own 7 s deadline.
The budget stops applying once response headers arrive. A vendor answering at 6.9 s and streaming a large product page used to be at risk of cancellation at the exact moment it started paying off; the body is now read to the end.

0.1.66 (2026-08-12)

Changed

The status qualifier that prefixed the title, both descriptions, the README heading, the OpenAPI title and the MCP resource is gone from every surface, by owner decision. The non-affiliation sentence and the trademark attribution stay and now carry that posture on their own.

0.1.65 (2026-08-12)

Changed

Same removal across the Store metadata and the published build. Shipped ahead of 0.1.66, which finished the job in the package manifest and the internal docs.

0.1.64 (2026-08-11)

Changed

asin accepts up to 2048 characters instead of 512. It doubles as a full product URL and Amazon's tracking tail grows without warning, so a legitimate URL was refusable for a reason the caller could not see. 2048 is what browsers and proxies already treat as safe.
The README describes the interactive OpenAPI view without naming a Console tab, so a platform rename cannot make the documentation wrong.

Fixed

The manifest ships buildTag: "canary". apify push uploads the manifest and it overwrites the remote value, so a manifest saying latest promoted straight to production on a bare push. Every build now lands on canary and production moves only by an explicit tag move.

0.1.61 (2026-08-11)

Added

Full output schemas. tools/list now describes every field of every result, including the objects inside products, reviews and offers, which were previously published as opaque objects. A client can plan against the schema instead of calling the tool to find out what comes back.
## Quick Start at the top of this README and on the Actor's Input tab, with the real endpoint and the two ways to authenticate.
OpenAPI description of the HTTP surface, rendered on the Standby tab: endpoints, request examples, both security schemes, and the 401, 403, 406 and 500 responses.
coverage in the amazon://marketplaces resource, stating per tool which storefronts it serves.

Changed

Tool descriptions are shorter. Field lists moved to the output schemas, where a model can read them structurally.
Every business object is now defined once and feeds both the internal validator and the published schema, so the two cannot drift.

0.1.57 (2026-08-11)

Fixed

Input hardening. amazon_bestsellers accepted a category containing path separators, which let a caller steer the request to another Amazon path. The category is now a single slug of letters, digits and hyphens, and the handler encodes it as well. Refused calls cost nothing: no fetch, no charge.
Tool arguments are validated against a closed schema. An unknown property is refused with -32602 instead of being dropped in silence, which previously meant a caller could be billed for a call whose argument was never read.
Length limits on query (256), category (100), marketplace (8) and asin (512), published in tools/list.
error.message is capped at 400 characters.
A transport failure now returns its eventId in error.data, so the identifier in the server log is one the caller can actually quote.

0.1.49 (2026-08-10)

Added

Managed-vendor fallback after both direct fetch tiers fail on a block or a rate limit, with a per-vendor circuit breaker. Billing still happens only after a fully validated result.

Changed

Fetch budget restructured to 40 s for the ladder and 5 s after it, split between the direct phase and the vendor phase, so a slow vendor cannot eat the time the parser needs.

0.1.34 (2026-08-09)

Changed

amazon_search withdrawn on FR. Five spaced searches on amazon.fr all came back as challenges, on both fetch tiers, while the product and best-seller pages answered normally through the same proxy country. The call now returns UNSUPPORTED_MARKETPLACE before any request is made, so it costs nothing.

0.1.30 (2026-08-08)

Fixed

Container kills under load. The V8 old-space limit sat above the container limit, so the kernel reclaimed the process before V8 escalated its collections.

Amazon MCP Server — Product Search for AI Agents

longanimous_bracken/amazon-scraper

MCP Server for Amazon product search. Search by keyword across marketplaces. AI-agent ready — returns prices, ratings, ASINs, and availability.

petteri mähönen

Amazon Trends MCP

trendsmcp/amazon-trends-mcp

Amazon product demand data as an MCP tool for AI agents. Weekly search volume, growth rates, and live Amazon best sellers - no Amazon API key needed.

TrendsMCP 25+ data sources, one connection

Amazon EU Product Search & Offers

truenorth/amazon-eu-product-search-offers

Scrape all prices in one request for the main stores in the European market (Amazon.de, Amazon.fr, Amazon.it, Amazon.es, Amazon.co.uk)

TrueNorth

Amazon Search Scraper

good-apis/amazon-search

Danny

Amazon Product Scraper

cerebral_aluminum/amazon-product-scraper

Benny

Amazon Product & Price Tracker — Monitor Prices

oneary/amazon-product-and-price-tracker-scraper

Track Amazon product prices, availability, rankings, and reviews across Amazon marketplaces.

Luan M.

Amazon Reviews Scraper

good-apis/amazon-reviews

Danny