Run your store's AI co-founder inside ChatGPT or Claude: no credits for the thinking.

What MCP is, in two plain sentences

MCP stands for Model Context Protocol. It is an agreed way for an AI assistant to use an outside tool, so ChatGPT or Claude can call your store instead of guessing about it.

SitesPlaced runs an MCP server at https://sitesplaced.com/api/mcp. You connect it once and sign in with your SitesPlaced account, and from then on the assistant you already pay for can read and change the store you already own.

Every tool calls the same server action the dashboard calls. The same ownership check, the same plan gate, the same credit ledger, the same record of what happened. None of this is a side door with looser rules.

Why the price changes when your assistant does the thinking

Here is the rule the code is written around: credits pay for our model calls, so when there is no model call of ours there is nothing to bill.

In the dashboard an SEO audit is 5 credits. That is fair. Our model reads your catalogue, judges it and writes the new copy, and we pay a model provider for that.

Over MCP the same job is split differently. get_seo_state hands your assistant the store's SEO title and description and every product's name, description and current meta description. Your assistant writes the new copy. apply_seo_copy saves it. No model of ours ran at either end, so both calls charge zero.

The exception is generate_store_image. That one really does bill an image model on our account, so it costs exactly what it costs in the dashboard: 2 credits for a draft, 12 for standard, 25 for premium, 45 for ultra.

One session, start to finish

Here is a Saturday morning with the connector added to Claude. Read it as a script: the words are yours, the tool calls are the assistant's, and the credit line is what it cost.

1. Read my store's numbers. The assistant calls get_store_insights. Back come orders in the last 30 days, best sellers by units and by revenue, and, when your plan includes analytics, unique visitors, conversion rate and the most-viewed products. 0 credits.
2. Tell me what to push this weekend. No tool call at all. The assistant reasons over the numbers it just read and argues for two products. In the dashboard this is the part you spend an agent message on. 0 credits.
3. Show me the SEO on every product. get_seo_state returns the store's SEO title and description and, for each product, its name, description and meta description as it stands today. Sixty products by default, up to two hundred if you ask. 0 credits.
4. Rewrite the six that need it. Your assistant writes them. Ask to see all six before anything is saved, and tell it to work only from what the product description already says. 0 credits.
5. Save it. apply_seo_copy writes them back. The store SEO title takes up to 70 characters, the store description up to 170, and up to 100 product meta descriptions go in a single call. 0 credits.
6. Put 10% off on from Friday evening to Sunday night. schedule_promotion sets a coupon that only works inside that window and an announcement bar that shows only while it runs. Give it ISO datetimes with the offset, like 2026-10-17T18:00:00+05:30. The store switches both on and off by itself. 0 credits.
7. Make one photo of the best seller on a marble table. generate_store_image bills a real image model. At standard quality that is 12 credits, and the image lands in the store's image library. 12 credits.

Total for the morning: 12 credits, out of the 1,500 the plan includes each month. The same work done in the dashboard would be an SEO audit at 5 credits, the image at 12, and a credit for every message to an agent. The difference in rupees is small. The difference in habit is not: you stop counting before you ask.

get_ai_credits is on the connector too, also free, so the assistant can check your balance and tell you what an image will cost before it offers to make one.

What the six tools cost

The credit prices for images are the published ones, the same in the dashboard, the same on the monthly and yearly plan, and the same for top-up credits. A generation that fails is refunded automatically, so you never pay for an image that did not arrive. The whole price table, including the actions that only exist in the dashboard, is in AI credits explained.

How to connect

1. ChatGPT. Settings, then Apps and Connectors, turn on Developer mode, Create connector, and paste https://sitesplaced.com/api/mcp.
2. Claude on claude.ai. Settings, then Connectors, then Add custom connector, and paste the same URL.
3. Claude Code. Run claude mcp add --transport http sitesplaced https://sitesplaced.com/api/mcp, then /mcp to sign in.
4. Cursor. Add { "mcpServers": { "sitesplaced": { "url": "https://sitesplaced.com/api/mcp" } } } to .cursor/mcp.json.

Authentication is OAuth. You sign in to your SitesPlaced account and approve access, so there is no API key to copy and keep safe, and you can revoke access from your account settings at any time. The connector only ever sees sites owned by the account you signed in with. Step-by-step notes and the rest of the tools are at /developers/mcp.

A good first thing to say, once it is connected: "list my sites, then read the insights for the store, and tell me which product to push this week and why." If the answer contains a number you recognise from your own dashboard, the connection is working.

What we did not expose, and why

Sending a broadcast is not a tool. Broadcasts are on the plan and they work: one send reaches your newsletter subscribers and your past buyers by email, and goes out on WhatsApp through a Meta-approved template once your Cloud API is connected. We left the sending out of MCP on purpose. It mails real people, and a tool call has no confirmation step a person actually sees. Your assistant can write the copy; you press send on a screen that shows how many people will receive it.

Also out of reach through this connector: billing, plan changes, custom domains, tracking scripts, and your customer and order records. get_store_insights returns counts, best sellers and rates. It does not return a list of your buyers.

And two things are not built at all, on the plan or over MCP: replies to WhatsApp messages that arrive on your own number, and AI video. We would rather you read that here than find out after paying.

The honest limit

An assistant is only as good as the numbers it read. A store three weeks old will get thin insights back, and an assistant asked for a strategy from thin data can fill the gap with a confident guess. One question protects you: "which number are you using?" If it cannot point at one that came from get_store_insights, it is guessing, and you should treat the answer as an opinion.

There is also no Apply card out here. In the dashboard, every write an agent proposes waits behind a card you tick. Over MCP it is the chat app that asks you, and chat apps differ in how clearly they ask. Read the tool call before you approve it, especially an apply_seo_copy with a hundred products in it and a schedule_promotion with a discount in it.

Then check the work. A meta description that adds a material, a size or a certification your product page never claimed is a returns problem later, not a copy problem now. A generated image needs its colour, its shape and any text on it checked before it goes near a live product page. The assistant is fast and it can still be wrong, so look before you publish.

FAQs