Amazon Seller Assistant plugin for Claude: add listing media

To use Amazon's Seller Assistant plugin with Claude for listing work, connect the plugin for your Amazon data and add Sume's hosted MCP server for the media. The plugin supplies what is weak on a listing; Sume supplies the images and clips, with a spend cap you set before anything is generated.

Amazon's announcement says the plugin connects a seller's Amazon data (listings, performance metrics, inventory, sales analytics) to outside AI tools so an agent can read from and act on the business. It launches first in Amazon Quick, is in beta with Anthropic's Claude, and is for sellers in Amazon's U.S. stores, with international expansion to follow (read 2026-10-04).

What does the plugin do, and what does it not do?

Amazon lists data access and actions such as restocking and pricing adjustments with audit trails, and says human approval is required before actions are executed. Amazon's page describes no image or video generation in the plugin, so the media step needs a second tool.

How do you connect Sume's hosted MCP server to Claude?

Sume's hosted MCP server lives at https://mcp.sume.com/mcp. In Claude Code, the quickstart gives two commands; interactive clients should use the OAuth flow rather than pasting an API key into chat.

claude mcp add --transport http sume https://mcp.sume.com/mcp
claude mcp login sume

Why do generation calls fail at first?

Hosted OAuth grants read-only access (mcp:read) by default. Paid tools such as generate_image return insufficient_scope until you turn Write on at consent, which adds mcp:write. There is no separate paid scope. Every write or paid call also needs an idempotency_key.

That gives you a second approval layer on top of Amazon's: Claude can read your listing gaps through the plugin, but Sume will not spend unless you granted Write.

What should you ask Claude to do?

A workable sequence, written as instructions to the agent:

• Read the plugin data and list the five listings with the thinnest media.
• Call tools_list and tools_schema for generate_image so the parameters are confirmed before use.
• Run each generation with dry_run set to true first, to see the admission and cost preview without submitting.
• Submit for real with max_spend_usd and a stable idempotency_key per listing, then wait on the jobs with jobs_wait.
• Show the resulting media URLs and wait for your approval before touching any listing.

How do you keep the batch bounded?

When a turn needs three or more calls of the same shape, such as one generate_image per listing, the hosted server's script_run tool runs a short program on the Sume side with limits on calls and paid calls. Set the paid-call limit to the number of listings you approved. Amazon's own page says nothing about Sume, and the plugin does not call it; you are wiring two connectors into one conversation.