SellerMate now has two ways to get your Amazon data out of SellerMate and into something else: the MCP server you connect to Claude or ChatGPT, and the Public API you call from code. They share the same engine, the same seventy-odd tools and the same guardrails, so the question we get most is simply which one to use.

The short answer: MCP when a person is in the loop, the API when software is. Here is the longer one.

What each one is

MCP is the Model Context Protocol, the standard AI assistants use to talk to outside tools. Add SellerMate to Claude from its directory, or paste one URL into ChatGPT, Cursor or Gemini CLI, and the assistant can call SellerMate's tools in the middle of a conversation. You ask why ACoS jumped last week; it pulls the campaign data, compares the weeks and answers. You say "negate those three terms"; it proposes the change and applies it within your guardrails.

The Public API is the same set of tools over plain HTTPS. Every tool is a POST to /public/v1/tools/{name} with a JSON body and an X-API-Key header. There is no conversation and no model in the middle unless you put one there. Your code calls it on a schedule, in a web app, from a spreadsheet script, wherever.

Both are included in every plan. Both are permission-checked, policy-guarded and audit-logged in the same way.

When MCP is the right choice

Someone is asking a question. Analysis, diagnosis, "what changed", "which campaigns should I look at". The assistant decides which tools to call based on the question. You never think about tool names.
You want one-off actions with a human approving them. Negating a list, pausing a campaign, shifting a budget, after reading the reasoning. MCP makes the proposal visible before it applies.
You don't want to manage keys. Claude's directory connection is one click and the grant follows you across claude.ai, Claude Desktop and Claude Code. Nothing to paste, nothing to store.
Your team lives in chat. Account managers who would never open a terminal will ask Claude a question. That is where the adoption is.

When the API is the right choice

It runs without you. A Monday report, a noon budget check, a nightly sheet refresh. Nobody is typing a prompt at 8am; a cron job is making the call.
Other people see the output. Client dashboards, Slack channels, emails, Looker Studio. The data has to land somewhere other than a chat window.
You need every row. An assistant summarises. A script walks all forty pages of search terms and writes them to a CSV.
You are building a product or an agent of your own. Your own assistant, your own rules engine, your own BI. The API is the layer underneath.
It has to be predictable. The same call returns the same shape every time. No model in the loop means no variation in what gets fetched.

The same guardrails either way

This is the part people ask about most, and the answer is reassuring. A write through MCP and a write through the API go through the same checks: your workspace's budget and bid limits, never-pause and never-archive lists, ACoS guards. A change that clears them applies. One that doesn't is blocked and returns a 422 with the rule that stopped it, and nothing reaches Amazon.

Keys and connections are scoped the same way too. An API key can be read-only and limited to specific Amazon accounts. An MCP connection carries the permissions of the user who made it. Both are revocable, and everything either of them does is in the audit log.

Using both, which most teams do

A common pattern:

Explore in MCP. Ask Claude what is going on in an account. Find the pattern worth automating.
Automate with the API. Once you know you want "every Monday, every client, these numbers", paste one of the prompts into Claude and let it write the script.
Keep asking in MCP. The script handles the routine; the conversation handles the exceptions.

The two even help each other. Claude knows the SellerMate tools from MCP, so when you ask it to write a script against the API it already understands what get_campaign_performance returns.

A quick decision guide

A person asks, a person reads the answer: MCP.
Runs on a schedule: API.
Output goes to a client, a channel or a sheet: API.
One action, approved by a human: MCP.
Thousands of rows: API.
Your own app or agent: API, possibly with MCP inside it.
You are not sure yet: MCP first, it is one click.

Getting started with either

For MCP, open the MCP page and add SellerMate to Claude from the directory, or copy the connector URL for other clients.

For the API, create a key from Profile → API Keys, then either read the documentation or skip straight to the prompt library and let your assistant read it for you.

Get your API key · MCP for Claude · Developer API

API or MCP? Which SellerMate integration to use for what