Amazon MCP servers compared (2026)

Updated August 19, 2026: added Hector MCP, which our original research sweep missed.

Your AI assistant is smart about everything except your Amazon business, because it cannot see your Amazon business. MCP (Model Context Protocol) fixed the plumbing problem: it is the open standard that lets an assistant like Claude or ChatGPT call real tools against real accounts. A year ago there was almost nothing to connect. As of August 2026 there are more than a dozen Amazon MCP servers, and they differ in almost every way that matters: what data they see, whether they can change your account, what stands between an AI's suggestion and a live change, and what they cost.

This is our map of the landscape - the options a seller would actually shortlist, verified firsthand. It is not exhaustive: the public MCP registries list thousands of servers and new Amazon ones appear monthly. Full disclosure up front: we make one of these. Ours appears last, the same rules apply to it, and where a competitor is better at something we say so. Every claim below carries an as-of date; if we got a fact about your product wrong, email us and we will correct it with a visible note.

Amazon's own servers

Amazon Ads MCP Server (closed beta November 2025, open beta February 2, 2026) is the gravity well the whole category orbits. It is hosted by Amazon, free to connect as far as Amazon's pages say, covers sponsored ads, DSP and AMC with 50+ tools, and its writes are real: create, update and delete campaigns, change budgets. Two things to know before you assume it is the obvious choice. First, it requires active Amazon Ads API credentials - a developer registration most sellers do not have. Second, nothing in the server enforces review of a change: Amazon's own announcement describes campaigns that "only need review and approval," which is a workflow suggestion, not a mechanism. It is also advertising only - no inventory, no margins, no organic data.

On the Seller Central side there is no hosted first-party MCP for seller-account data as of mid-August 2026. What exists is a local developer-assistant package Amazon labels "educational examples," aimed at people building SP-API integrations, and the in-console Seller Assistant, which external AI clients cannot drive. If you want your assistant to see orders, settlements or inventory, you need someone other than Amazon.

The established seller suites

Helium 10 MCP (July 16, 2026) has the widest data span of the big suites: keywords, reverse-ASIN and product research, brand analytics, ads performance, profit and inventory, with writes on bids and budgets behind a stated approval step. If you already live in Helium 10's research stack, this is a genuinely strong option. The catch is the gate: MCP is included only on the Diamond plan, $279 to $359 a month as of August 2026, and its pages publish no refresh cadence or rate limits.

Jungle Scout MCP (July 8, 2026) is a different animal: read-only market intelligence - market share, share of voice, category growth, competitor brands - drawn from a panel Jungle Scout describes as 18% of Amazon's GMV. For "what is happening in my category" questions it has depth nobody else here matches. It is not a seller-operations tool (no P&L, no PPC management), and it is gated to the enterprise Cobalt tier behind a sales demo.

Pacvue MCP Server (May 14, 2026) is deliberately read-only: it pulls Pacvue's reports across thirteen retail media networks, which makes it the only multi-retailer option in this list. Writes live in the separate Pacvue Agent product with approvals and an audit trail - mature enterprise governance, but in Pacvue's platform and Slack, not over MCP. Enterprise pricing, existing customers only.

Seller Labs MCP (November 5, 2025 - the earliest mover in the whole category) covers both seller data and ads from its Data Hub, and is by far the cheapest suite entry: free under a revenue floor, revenue-tiered above it, 14-day trial. Campaign writes shipped later with "guardrails" such as budget warnings - which is real protection, but note the difference between a warning and an enforced approval step.

Two names you might expect are absent by choice. Sellerboard has no MCP or AI surface at all as of August 2026 - and remains the cheapest polished profit analytics around ($19 to $79 a month). Perpetua made the opposite architectural bet: a proven autonomous bidding engine with no conversational access. If you want goal-based hands-off optimization rather than an AI you talk to, that is a coherent choice - it is simply a different product category.

The AI-native services

DataDoe (launched May 2026, built by the Deltologic team) takes a warehouse approach: SP-API, Ads API and - uniquely here - Vendor Central, normalized into 100+ tables refreshed daily, with BigQuery access for analysts. Writes are disabled by default and enabled per action type, with dry-run validation, and it has the best-documented security posture of the independents: annual Amazon PII audit, third-party pen testing. Its pricing is a $97/month base that meters three dimensions - 10 million data rows, a monthly allowance of AI actions, one seat - with unlimited account connections included, which agencies will notice. Growth past the included capacity comes as add-ons (additional row blocks, seats and actions), so a larger catalog should price those in rather than the sticker alone. Most tables are daily refresh, so it is not the tool for intraday questions.

AgentCentral ($39 to $199 a month by order volume) has the broadest surface of the hosted wrappers - ads, orders, finance, inventory, catalog - with preview-and-confirm writes and a five-minute setup that needs no developer credentials. Two honest cautions: its published tool counts disagree with each other across its own pages, and we could not find a legal entity name or any security certification behind the service you would be granting Amazon account access to.

Adspirer is the cross-platform play: one MCP for Google, Meta, TikTok, LinkedIn and Amazon Ads together, with a real free tier and a sane default that new campaigns arrive paused. Amazon is a minority of its surface (61 of 400+ tools) and there is no Seller Central data at all - right tool if ads across many platforms is your actual problem.

Hector MCP ($300/month, by Hector Ai Global) is an ads-only intelligence layer that positions itself explicitly as a complement to Amazon's first-party server rather than a competitor to it. Its genuinely clever piece is the bulk-operation architecture: one instruction ("lower bids 15% on everything below my target") executes through backend bulk APIs instead of the AI making hundreds of individual tool calls and drowning its own context window. It supports Seller and - unusual at this price - Vendor accounts, with unlimited profiles, which is why its pitch leans agency. Know two things going in: it is advertising only (DSP and AMC are listed as coming soon, no seller financials), and as of August 2026 its pages describe direct execution from conversation and autonomous agents with no approval step described anywhere - the exact opposite end of the write-safety spectrum from a staged-approval model. It also requires a paid ChatGPT or Claude subscription on top.

MarketplaceAdPros (AI Connect from $10/week) is the cheapest paid on-ramp to Amazon Ads data in an assistant, and PPC Prophet offers a free read-only tier (100 calls/day) with opinionated diagnostic tools, limited to a 90-day data window. Both are ads-only and service-dependent.

Open source, for the credentialed

KuudoAI's amazon-ads-mcp is the most substantial genuinely open codebase (MIT, 337 commits as of August 2026): campaigns, reporting, DSP and AMC across NA/EU/FE regions, reads and writes, self-hosted via Docker. jay-trivedi's amazon-sp-mcp (MIT) is a clean read-only SP-API server: orders, sales, returns, inventory, 16 tools. Both require what the hosted services exist to spare you: your own registered developer credentials. If you have them and want no third party in the loop, this corner of the map is honest, free and yours to run.

The comparison, in one table

Where AMZ Vault sits, and what we cannot do

Our MCP server exposes 101 tools over the same warehouse our own dashboards read: settlement-accurate profit down to the fee level, all three sponsored ad types, search terms and keywords, inventory, brand analytics and forecasting - and every write the AI proposes is staged as a reviewable diff that executes only after you confirm it. That approval gate is enforced by the architecture, not by a prompt, which as far as we found is unique in this list. It is included in the $99 flat plan, not an upper tier.

What we do not do: product research or competitor discovery databases (Helium 10 and Jungle Scout are better there and we use such tools ourselves), anything outside Amazon (Pacvue and Adspirer own multi-platform), autonomous execution even if you want it (a deliberate constraint), and multi-account on the flat plan (that is a conversation, not a checkout). We are also the newer product in this list - the open demo exists so you can judge that trade for yourself without a signup.

However you choose: prefer the tool whose failure mode you can live with. Data you can read is easy to shop for. Writes are where the products genuinely differ - autonomous by design, a warning, a convention, a default-off switch and an enforced approval gate are five different answers to the same question, and it is your account that carries the answer.

Facts verified August 18-19, 2026 from vendor pages, official Amazon announcements and public repositories; dated source links inline. Related: how our AI connection works · try the whole platform with no signup