Setting up an Amazon agency workspace: plan, client connections, team, reports and your own AI, step by step

An agency that runs several Amazon accounts does not have a dashboard problem. It has a "which of my clients needs me today" problem, and no per-account view answers it. This guide walks through the agency setup in AgentXray from the plan to the first client report, in the order the product asks for it. Every step describes what the product does today; the last section lists what it does not do yet.

What you get at the end

One list across every connected client account, ranked worst first. Each row names the account it belongs to and the source its number came from: advertising search terms, inventory, the listing's own health issues in Amazon's wording, margin from settlement line items, or the account health tier Amazon itself reports. The list is rebuilt every hour. Your own Claude or ChatGPT can read the same accounts and draft changes; a person approves each one before it reaches Amazon.

Step 1: the plan and the trial

Agency Starter is $99 a month for up to 3 connected client accounts and 3 team seats (the owner plus two). It starts with a 14-day trial that needs no card. Add a card under Billing before day 15 to keep the workspace open; if no card is on file by then, the subscription is cancelled and nothing is charged. Larger agency plans, for 10 and 50 client accounts, are arranged by conversation rather than bought from the pricing page.

Pick Agency Starter on the pricing page. The checkout has no card step. When it returns, the agency navigation is on: a Clients tab, an All clients view and the client report.

Step 2: connect the first client

An agency cannot connect a client's Seller Central itself, because Amazon asks for the client's own login. So the workspace owner makes a one-time invite link in the Clients tab: the client's name, and optionally their email and store. AgentXray can email the link, or you paste it into the conversation you already have with the client.

The link opens a page that names your agency. The client picks their store and approves on Amazon's own consent page, with their own login. They need no AgentXray account, and AgentXray never sees their password. The link works once and for 7 days, and you can withdraw it before it is used. An Amazon account that is already connected to another AgentXray workspace is refused, not moved. The fourth account on Agency Starter is refused as well; nothing already connected is ever disconnected by the ceiling.

The owner can also connect an account directly, the same way a single seller does, when the agency itself holds the login.

Step 3: name the client

Give each connected account a label. The label is what the All clients list, the client report and your AI connector show, so a label that matches the name your team uses in conversation saves a lookup every time.

Step 4: seats and roles

Three roles exist. A viewer reads everything and changes nothing. An operator does the daily work: works the list, drafts changes and approves them. The owner can also connect and remove client accounts and manage the team and billing. Invite team members from the Team panel by email; an invitation that has not been accepted yet still counts against the seat ceiling, so a seat is never promised twice. People who belong to more than one workspace switch between them from the workspace switcher.

Step 5: work the list

The All clients view is the agency's morning. Every open action across every account, worst first, each with its account, its source and its reason. Decide on each row from there. The client list beside it shows, per client, live products, SKUs sold and profit for the last 7, 30 or 90 days, in the client's own currency. Currencies are never added together or converted, and a client with no sales says so instead of showing zero. Profit per client is before advertising, for a reason explained in the last section.

Step 6: approvals

Every change an AI drafts, whether a bid, a budget, a negative keyword, a listing update or a price, waits in one approval queue for the workspace. Operators and owners approve from the browser. The connector token your AI holds cannot approve: the approval endpoint refuses it. A leaked connector can propose anything and change nothing.

Step 7: the client report

For each client, a report of what was found, what was changed and what each change did, for the last 7 to 180 days. Each change carries its own result and is never summed into one number; a change that has not settled yet says the date it will be judged. Where data is missing, the report names the gap instead of filling it. The report is plain text, copied to the clipboard, written for the email thread you already have with that client. Viewers can read it too.

Step 8: your own Claude or ChatGPT

One connector covers the whole workspace. Ask it to list the connected accounts and it names each one with its marketplaces. Any question about a marketplace takes an account argument; where more than one connected account could match, it refuses to guess and asks which client you meant. Drafts it files land in the same approval queue as everything else. The connection steps are in the MCP guide, and the MCP servers for Amazon sellers round-up compares this connector with Amazon's own and others.

What is not included today

• Client logins. The client approves access on Amazon and reads the report you send; there is no branded workspace the client signs into.
• White-label or branding of any kind.
• Ad spend per client. Amazon Ads is connected per workspace, not per client, so each client's profit is shown before advertising rather than with the spend divided between clients by a guess.
• A PDF or CSV of the client report. It is plain text, copied to the clipboard.
• A single AI tool that returns the cross-client list. The connector reads accounts one at a time.
• Self-serve checkout for the 10- and 50-account plans.
• Anything that runs unattended. Every change is approved by a person first.

Where to go next

• For agencies: the agency page.
• Agency pricing.
• Triage across client accounts: why one list beats fifty dashboards.
• Security: how seller data is kept apart between workspaces.