Amazon 卖家如何用 MCP 连接 Claude 和 ChatGPT

Every week another Amazon seller asks some version of the same question: can I just connect my account to ChatGPT or Claude and ask it things? The answer in September 2026 is yes — and the interesting questions are the three hiding behind it. Through what? With what permissions? And what happens the first time the AI wants to change something instead of just reading it? The 'through what' is MCP, the Model Context Protocol — the open standard that has quietly become the way AI assistants plug into outside systems, with more than 10,000 public servers by the time Anthropic handed its governance to the vendor-neutral Agentic AI Foundation in December 2025. Amazon itself has stepped in twice: its Ads MCP server, in open beta since February 2, 2026, covers advertising, and on September 23, 2026 it added the Selling Partner plugin, which brings Seller Assistant's view of your listings, inventory, and sales into Claude in beta. Anything beyond what those two expose still runs through third-party servers. I managed 57 Amazon accounts before building SellerForge (disclosed up front: SellerForge is our product, and it includes an MCP connector). This guide is the one I wish existed when I first wired an account into a chat window: what MCP actually is in seller terms, what Amazon officially supports, the four ways to connect, and the safety ladder that should decide which one you use. ## What Is an MCP Server for Amazon Sellers? An MCP server is a translator that sits between your AI assistant and your Amazon business: it holds an authorized connection to your seller data and exposes it as tools the AI can call. Ask Claude 'which SKUs run out inside 30 days?' and instead of guessing, the AI calls the server, gets your own numbers, and answers with them. The protocol underneath is worth thirty seconds of history, because it explains why this suddenly works everywhere. Anthropic open-sourced the Model Context Protocol in November 2024 as a universal way for AI models to reach external tools — one standard instead of a custom integration per app. OpenAI adopted it across its products in March 2025, Google and Microsoft followed, and in December 2025 Anthropic donated the protocol's governance to the Agentic AI Foundation. By then MCP had over 97 million monthly SDK downloads and more than 10,000 active public servers. It is the de facto port that every serious AI client now speaks: Claude, ChatGPT, Cursor, Claude Code, Gemini, Copilot. ## Does Amazon Have an Official MCP Server? Yes, and since September 23, 2026 there are two official options. The first covers advertising only: the Amazon Ads MCP Server entered open beta on February 2, 2026, with roughly 50 tools spanning Sponsored Products, Sponsored Brands, Sponsored Display, DSP, and AMC, available to anyone holding Amazon Ads API credentials. Until Accelerate 2026 there was no official MCP route for Seller Central data at all — none for orders, inventory, fees, returns, or account health. Two things follow from that. First, if your only goal is ads automation and you have developer credentials, the official Ads MCP is genuinely the right tool — it is free with API access, first-party, and will only deepen. Second, check what any 'official Amazon MCP' claim actually covers: Amazon's own options are the Ads server and the Selling Partner plugin below, and anything beyond what they expose runs through SP-API under a registered developer, or through a platform that already holds your authorized data. Amazon Accelerate 2026 (September 22–24, Seattle) made agentic AI its headline track and delivered the Seller Central half. Claude's connector directory lists the result as Amazon's Selling Partner MCP connector (checked September 25, 2026). Per Amazon's announcement, it is in beta for sellers in Amazon's U.S. stores, reads listings, inventory, sales, and performance metrics, and can act on your account, for example adjusting a price, after you review and approve each action. ## Is It Safe to Connect Your Amazon Account to an AI? It is safe when the connection runs through Amazon's official APIs under OAuth, is scoped and revocable, and puts a human approval step in front of any write. It is dangerous when an agent holds your password or writes to your account with no gate. The entire safety question compresses into one distinction: authorized API access versus credential-holding agents. But 'uses the official API' is only half the safety story. The other half is what the AI is allowed to do once connected, and this is where I want to give you the frame I now use for every connector I evaluate — ours included. [The Read→Propose→Apply Ladder: every AI-to-Amazon connection sits on exactly one rung. READ — the AI can look and answer, nothing more. PROPOSE — the AI can stage a change that a human reviews and approves. APPLY — the AI executes a change a human already approved, inside caps and an audit trail. Know the rung before you connect, and never hand an AI rung three through a connector that never built rung two.] Run every option you are considering up that ladder. A read-only reporting pipe is rung one by design — safe, and limited. A raw SP-API bridge that exposes write endpoints jumps straight to rung three with no rung two: the write goes through the moment the model decides to make it, which is exactly what you do not want an unsupervised language model doing to live listings or campaigns. Across the 57 accounts I managed, the expensive mistakes were rarely bad analysis — they were good analysis executed without a second pair of eyes. The approval step is not bureaucracy; it is the entire difference between an assistant and a liability. • Insist on OAuth through official APIs — any tool that asks for your Seller Central password is disqualified, whatever else it does. • Confirm you can revoke access in one place, and that revocation actually kills the token rather than just hiding a menu item. • Ask where writes go: straight through (rung three with no gate), into an approval queue (rung two), or nowhere (rung one). • Demand an audit trail — who proposed what, when, why, and who approved it. If the vendor cannot show you one, assume there is none. ## The Four Ways to Connect Seller Central to an AI Four routes exist in 2026, and honest routing matters more than brand loyalty: they solve different jobs, and two of them are not even competitors to the others. Here is the whole landscape in one table, including where each route genuinely wins. ## How to Connect Claude or ChatGPT to Seller Central The connection takes about five minutes in any MCP-capable client, and the shape is identical everywhere: point the client at a server URL, authorize, ask. Here is the guarded route, using the SellerForge connector as the example — the same steps apply to any hosted MCP server, with its own URL. 1. Pick your rung first. Decide whether this connection should read, propose, or apply before you authorize anything — not after the AI surprises you. 2. In Claude (web or desktop): open Settings → Connectors, choose to add a custom connector, and paste the server URL: https://www.sellerforge.ai/api/mcp 3. In ChatGPT: open Settings → Connectors, enable developer mode, create a new connector, and paste the same URL as the server endpoint. 4. In Claude Code, it is one terminal command: claude mcp add --transport http sellerforge https://www.sellerforge.ai/api/mcp — Cursor takes the same URL in its mcp.json. 5. Authorize your account when prompted. You sign in to SellerForge; the AI never sees a password, and Seller Central credentials are never involved. 6. Ask a question you already know the answer to — your current Account Health Rating or order defect rate — and confirm the numbers match Seller Central before you trust anything else it says. That last step is an old operator habit that transfers perfectly: calibrate the instrument before you fly it. Reading works on every SellerForge plan including the free trial, so the calibration run costs nothing. The connector is read-only today; guarded proposals return on paid plans as SellerForge enables them. Revoking a client takes one action in Settings → Connections. ## What Should You Actually Ask Once Connected? The highest-value questions are the ones you currently answer with a 40-minute export-and-pivot ritual, because those are the ones you have quietly stopped asking. Connected over MCP, they cost one sentence each. These are the six I run weekly, and they map to the six data groups a whole-account connector should expose. • Profit: "What was my true net profit last month after fees, ads, and COGS — and which three SKUs dragged it down?" • Advertising: "Which campaigns have the worst ACoS this week, and where is spend being wasted?" — then sanity-check the answer against your ads dashboard. • Keywords: "Which search terms are bleeding money, and which are worth harvesting to exact match?" • Inventory: "Which SKUs run out inside 30 days at current sell-through, and what should I reorder this week?" • Account health: "Is anything trending toward a threshold — ODR, late shipment, valid tracking — that I should fix before Amazon notices?" • Voice of Customer: "What are customers actually complaining about in recent reviews and returns, by ASIN?" A disclosure on our own connector, checked September 25, 2026: SellerForge's MCP answers the account-health question today, lists inventory on hand, and returns Voice of Customer on Growth and above. It cannot answer the profit, ad-waste, and search-term questions yet: SellerForge's Profit & Loss module is temporarily withheld, and Amazon Ads performance ratios and waste review are paused while the underlying data is verified, so the connector leaves those tools out rather than guess. Restock advice appears only where the inventory and sales history qualifies. One honest caution from someone who has done this daily for a year: a connected AI is confidently wrong less often, but it is not never wrong. Timezones, attribution windows, and half-loaded data will occasionally produce a plausible bad number. Keep the calibration habit — spot-check anything that would change a real decision. Rung two exists precisely so that an occasional bad number produces a rejected proposal instead of a paused campaign. ## Where SellerForge Fits (and Where It Does Not) SellerForge is our product, so weigh this section accordingly. The MCP Connector is included with every plan — no add-on fee — and plans run from $49/mo (Core $49, Growth $99, Pro $199, Agency $499, 7-day trial). Reading your account works on every plan, including the trial. The connector is read-only today; guarded proposals return on paid plans as SellerForge enables them, and when they do, every proposal lands in an approval inbox with the exact edit and the AI's reasoning. Caps, roles, and an audit trail stay in force no matter what the AI asks for; MCP cannot change its own permissions or route around policy. The direction of travel is not subtle. Adobe measured AI-referred traffic to US retail sites up 393% year over year in Q1 2026, and by July those visitors were converting 60% better than non-AI traffic. Shopping is going agentic on the demand side; operations are going agentic on the supply side; and the sellers who wire their accounts into their assistants now — on the right rung of the ladder — are building the reflexes everyone else will be scrambling to learn in 2027.