Start free trial

Model Context Protocol

Your Amazon account, as a tool your AI can call.

ScaleSKUs runs an MCP server over OAuth 2.1. Point any MCP-capable client at it and your AI can query the account, reason about it, and propose changes you approve.

Connect in two minutes Browse the tools

~/.config/claude/mcp.json

{
  "mcpServers": {
    "scaleskus": {
      "url": "https://scaleskus.com/mcp",
      "transport": "http",
      "auth": "oauth2"
    }
  }
}

No API keys. OAuth handles the handshake.

Why a protocol, not an integration.

An integration ties your data to one vendor's interface. A protocol lets any client call the same tools, so the choice of AI stays yours and changing it costs nothing.

MCP is an open standard. We implement the server; your client does the rest.

Client independence

Claude today, something else next year. The tools do not change, so neither does your setup.

No data movement

Nothing is uploaded into a model. The client calls a tool, the server computes, a result comes back.

Computed, not summarised

Tools return finished figures rather than raw rows, so the model has nothing to arithmetic its way through.

Your own agents

Anything that speaks MCP works, including internal tooling you build yourself.

What happens on one question.

A single turn, end to end. Note where the computation happens — the model never sees raw rows it has to summarise.

01 user

The question

“Why did sales drop last week?” typed into whichever client you use.

02 client → server

Tool selection

The client picks diagnose_sales_change from the advertised surface and calls it with a date range.

03 server

Scoped query

OAuth identifies the organization. Row-level security limits the query to your profiles before it runs.

04 server

Computation

The decomposition runs in the intelligence layer. Price, stock, traffic, conversion and ads are separated.

05 server → client

Grounded result

Finished figures with their sources return. If a figure is unavailable it is absent, not estimated.

06 model → you

The answer

The model explains the result and may propose an action. Applying it needs your approval and 2FA.

The tool surface.

Read tools return computed figures. Write tools return a proposal, never a completed change.

Tool	Returns	Scope
get_account_summary	Sales, spend, TACoS, margin and order counts for a date range, with period comparison.	Read
get_campaign_performance	Campaign rows with spend, ad sales, ACoS, budget-cap time and verdict against your target.	Read
get_search_terms	Search-term rows with clicks, orders, spend and a zero-order flag. Filterable by threshold.	Read
get_asin_performance	Per-ASIN organic and ad sales merged, with conversion trend and stock cover.	Read
get_inventory_risk	ASINs at or near stockout, plus any live campaigns still spending against them.	Read
get_search_query_performance	SQP funnel by term — impressions, clicks, cart adds, purchases — with your brand share at each stage.	Read
get_profit_and_loss	Revenue, fees, COGS, ad spend and net margin by period. Requires COGS to be set.	Read
get_plan_vs_done	Agreed actions from the last review against the change log, with status and value at stake.	Read
get_change_history	Every change made to the account, who made it, when, and the previous value.	Read
diagnose_sales_change	A decomposition of a sales movement across price, stock, traffic, conversion and advertising.	Read

Read tools cannot modify anything. Each returns computed figures with the Amazon report they came from.

A write, in full.

The model can propose. It cannot commit. Approval is a separate, human, 2FA-gated step outside the model's reach.

tools/call Model proposes

{
  "name": "propose_budget_change",
  "arguments": {
    "campaign_ids": ["c_8812", "c_8813"],
    "daily_budget": 2100,
    "reason": "capped 2:10pm, ACoS 18.2%"
  }
}

result Pending, not applied

{
  "status": "awaiting_approval",
  "proposal_id": "p_4f21a9",
  "affects": 2,
  "from": 1200, "to": 2100,
  "requires": ["human_approval", "2fa"],
  "reversible": true
}

Worth stating plainly

Amazon issues no read-only credential, so the authorization you grant is full access by design. That is the API, not our choice. What we control is our behaviour with it: the MCP surface exposes no tool that writes to Amazon directly, approval happens outside the model's reach, and every applied change is logged with its previous value.

See it connect.

A short screen recording of the whole connect flow in each client — every click, start to finish. Silent; pick a client and press play. The written steps are just below.

No audio — use the tabs above to switch client. The steps below spell out the same flow in text.

Connect your client.

Claude Native

Add the server in Settings → Connectors, or drop the config above into your mcp.json. OAuth opens in the browser.

ChatGPT Native

Add scaleskus.com/mcp as a connector in a Team or Enterprise workspace. Same OAuth flow, same tool surface.

Gemini Native

Supported via MCP in the Gemini app and Vertex AI. Register scaleskus.com/mcp and authorize once.

Your own agent Any MCP client

Any client implementing the MCP spec can connect to scaleskus.com/mcp. The tool surface is discovered, not hard-coded.

Also available

scaleskus cli

The same tools from a terminal. Useful for scripted checks and CI reporting.

REST API

For services that do not speak MCP. Identical scoping and approval rules apply.

Webhooks

Push notification when something moves — a stockout, a budget cap, a position loss.

Prompt library

Connected? Here’s what to ask.

220+ ready-to-use prompts, from a first plain-English question to advanced optimisation — filter by role and difficulty, then copy and paste.

Browse the prompt library

Technical questions.

Does the model see my raw account data?

It sees the computed result of a tool call, not the underlying rows. That is deliberate — a model asked to sum 40,000 rows will make arithmetic errors, so the arithmetic happens server-side.

What stops it acting without me?

No tool in the surface writes to Amazon. Write tools return a proposal object, and approval happens in the app behind 2FA, outside anything the model can call.

How is access scoped?

OAuth identifies the organization, and row-level security constrains every query to your profiles at the database level. A bug cannot leak another customer’s figures.

Can I limit which tools are available?

Yes. Tool access is per-seat, so an analyst seat can be given read tools only while an owner seat keeps the proposal tools.

What if I revoke access?

Sync stops immediately and the MCP endpoint returns unauthorized on the next call. Revoke from your account page or from Amazon directly.

What are the rate limits?

Amazon’s limits are the real constraint, not ours. Results are cached where the underlying report only refreshes daily, and every response states its freshness.

Point your AI at your account.

Thirty days free. OAuth takes about two minutes; the first full sync takes six hours.

Start free — 30 days