AI connectors

Connect Claude or ChatGPT and ask questions in plain language. The assistant reads out of the database in your browser, per question, through a connection you granted and can revoke — and anything it proposes waits in the Publish queue for you.

Before you start

Three things have to already be true, in this order:

1. The extension is installed and the first sync has finished. The data the assistant reads lives there, not on our servers.
2. The browser is connected to your MerchPPC account — the sign-in the extension asks for after its first sync. That authenticated connection is exactly what the assistant reaches your data through.
3. Chrome is open while you ask. Answers come from the database in your browser; there is no cached copy anywhere else to answer from.

Connectors are free during the beta. Claude and ChatGPT are separate connections: you grant and revoke each independently, and consenting to one never grants the other.

Set up Claude

Works on claude.ai and in Claude Desktop — both use the same connector, added by URL.

1. In Claude, open Settings → Connectors → Add custom connector and paste https://mcp.merchppc.com/mcp.
2. Approve it once, in your browser. Claude opens the MerchPPC consent screen. It shows you what Claude asked for — read-only, or read and propose — and you approve or deny. A read-only grant never implies the other.
3. Ask something. That is the whole of it. The grant belongs to your account; which browser answers is your primary-device choice on the account page, changeable without re-authorising.

Set up ChatGPT

Same connector, same consent screen, reached through ChatGPT's own connector settings.

1. In ChatGPT, open Settings → Connectors — on some plans this lives under Advanced → Developer mode — choose Add connector, and paste the same URL.
2. Approve it on the MerchPPC consent screen, as its own separate grant.
3. Enable it in the chat: pick MerchPPC from the tools menu in a conversation, then ask.

The exact address is printed on your own Account → Assistants page. If the two ever disagree, that page is the authority — it reads the value the server actually enforces, and a connector pointed at any other host is refused.

What to ask

• Give me a status check on my MerchPPC data.
• Which of my products are losing money on ads right now?
• What is my ACoS in the US this month, and is that good?
• Queue a negative for every search term that spent $5 with no order.

Behind the questions sit a dozen tools: account status, an account summary, products, a single product whole, a day-by-day series, coverage, the advertising structure, findings that already carry their judgement, a glossary, a diagnosis of the extension itself, and a way to send us feedback.

Exactly one of them proposes. It stages an advertising change into the Publish queue, where you approve the row before anything reaches Amazon — a bid or a budget moved either way, a pause on a target, an ad, an ad group or a campaign, an excluded search term, or a converting search term promoted to a keyword you bid on. Anything that would increase spend needs Pro; everything that lowers it works on the free plan, and a proposal held back for the plan is still shown to you with its real numbers. It cannot author a number: it names an action the extension already computed, and there is no argument anywhere in it that carries a bid, a budget or a keyword of its own.

It cannot create a campaign, split one, un-exclude a term, add a product to an ad group, or touch a price, a listing or artwork. Those stay yours.

Three honesty rules every answer follows

Money is never summed across marketplaces. They settle in different currencies and nothing normalises them. Omitting a marketplace partitions an answer; it never merges it.

Per-product advertising data only exists for roughly the last 90 days. Over a longer window an ad spend of zero means not known, not no ads ran. Sales figures are unaffected.

A database that has never published summaries says so, instead of reporting zeroes as though they were facts.

Every answer that reports data carries a basis block naming what the data could not see.

Permissions, limits, revoking

• Revoke any time at Account → Assistants. Each assistant is its own grant with its own revocation, and revoking is immediate.
• Quota: 30 calls a minute and 300 a day, shared across assistants — enough for a working session, not for a scraper.
• Devices: your plan decides how many connected browsers may answer. During the beta that is three; choose which one answers at Account → Browsers.
• Audit: every call is logged against the grant, so what did it read? has an answer. The record is metadata — when, which tool, how big — never the question or the answer.
• The relay stores nothing. Questions and answers pass through while you use it; your Merch and Ads data is never uploaded or cached on our side.

When something is wrong

Every message the connector produces carries its next step. The ones worth explaining twice are in Troubleshooting.