An AI agent is good at the slow part of e-commerce research: running ten searches, opening the results, comparing them and writing up what they have in common. What it lacks on its own is data. Ask a chat assistant which stores are growing in a niche and it can only guess from what it learned in training.

MCP fixes that by giving the agent tools it can call for live data. Connect an agent to the SellWinners MCP server and it can search Shopify stores, ads and marketing emails itself, then hand you a table or a brief. Here is how MCP works, how to connect your agent, prompts that get good results, and how the calls count against your plan.

What MCP is, in plain words

MCP stands for Model Context Protocol. It is an open standard for connecting AI assistants to outside tools and data.

It has two sides. An MCP server publishes a list of tools: the name of each one, what it does and the arguments it takes. An MCP client is the app you talk to, such as Claude Desktop, Cursor or VS Code. The client shows the tool list to the model. When your question needs data, the model picks a tool and fills in the arguments, the client runs the call, and the result comes back for the model to read before it answers or calls the next tool.

For research this changes the work. You ask a question in plain language, the agent chooses and chains the tools, and you review the answer. You don't write code, and the agent works with the same data as the SellWinners web app and REST API.

What the agent can do with SellWinners

The server exposes the product as ready-made tools. These are the ones you will use most for research:

ToolWhat it does
lookupTurns a brand name, a shop domain or a Facebook page id into ids the other tools use
search_shopsSearches or browses Shopify stores with filters
find_winning_productsProducts selling now in a niche
find_similar_shopsThe competitors of a shop, by id or domain
brief_competitorA compact brief on one competitor
search_adsSearches Meta ads by words in the copy, brand or landing page
scan_adThe full record of one Meta ad
creative_inspiration_packCreative inspiration for a niche
search_google_ads_librarySearches Google ads
search_tiktok_librarySearches TikTok ads and organic posts
search_emailsSearches brands' marketing emails
analyze_shop_emailsA shop's email strategy
analyze_brand_changesWhat changed for a tracked brand over a period
usage_getYour plan, monthly credits and how many you have used

There are more for tracking brands, saving items to boards and favorites, and sharing. The MCP page lists every tool with its arguments.

How to connect your agent to SellWinners

  1. Create an API key. Sign up, then create a key in your dashboard. Free accounts make keys too: the Free plan includes the web app and the MCP server (150 credits a month), without the API (see the plans and their prices). Create a key just for your agent, so you can revoke it on its own later.
  2. Add the server to your client. Your client needs three things: the endpoint URL shown on the MCP page, the streamable HTTP transport, and your key in an X-API-Key header (or Authorization: Bearer followed by the key). The MCP page has ready-made configuration for Cursor, VS Code and Claude Desktop. Claude Desktop connects through the mcp-remote bridge, which needs Node.js.
  3. Check that it works. Ask your agent: "How many of my SellWinners credits have I used this month?" It should call usage_get or check_credits, and neither costs a credit.

Other clients and agent frameworks work when they support remote MCP servers over streamable HTTP with a custom header. If you use another assistant, ChatGPT for example, check whether your version lets you add a remote MCP server with a custom header. If it does, the same three values are all it needs.

Prompts that work

Agents do best with specific questions: name the niche, the market, how many results you want and the shape of the answer. You can name tools to steer the agent, or leave the choice to it.

Product research

Use find_winning_products for "dog grooming". Then use search_shops to find Shopify stores in that niche selling to the US with at least 20,000 monthly visits, sorted by 30-day traffic growth. Give me a table of the top 10 with store, monthly visits, growth, best seller and its price.

Follow up with "for the top three, use brief_competitor and tell me what they have in common". Before you act on a candidate winning product, run your own costs at the market price through the profit margin calculator.

Competitor ad research

Use search_ads to list the active Meta ads of [competitor domain] that have run for at least 30 days, sorted by reach. For each one, give the opening hook, the headline, the call to action, the landing page and the days running. Then summarize the three angles it repeats most.

This turns an ad library search into a creative brief. Ask for the same brand with search_tiktok_library and search_google_ads_library to compare channels, or use creative_inspiration_pack for a whole niche when you need fresh hooks.

Store teardowns

Give me a brief on [competitor domain] with brief_competitor. Then use find_similar_shops sorted by monthly visits and compare the top five in a table: monthly visits, 30-day growth, active ads, product count. Finish with analyze_shop_emails for the first store: how often it emails, on which days and with which promotions.

The guide on how to analyze any Shopify store lists everything worth checking in a teardown.

Keeping watch

Add [competitor domain] to my brand tracker.

Then, once a week, ask:

Use analyze_brand_changes on [competitor domain] for the last 7 days. Which new ads did it launch, and which ads gained rank?

Tracking the brands you choose is part of the paid plans. The agent can also save what it finds: ask it to create a board and add the five best ads with create_board and add_board_item.

How tool calls use your credits

  • A tool call costs the results it returns. 1 credit = 1 result returned: one brand's details, one ad, or one row of a list. An empty result is free. MCP calls come out of your plan's monthly credits, the same credits the REST API uses.
  • Some calls are free. Failed calls cost nothing, and neither do usage_get, check_credits, lookup and lookup_filter_ids. Rate-limited calls and calls with rejected parameters cost nothing either.
  • One question can mean several calls. The teardown prompt above makes at least three: brief_competitor, find_similar_shops and analyze_shop_emails. Each costs the results it returns, and a lookup first costs nothing. Ask for fewer results to spend less.
  • The web app doesn't count. Browsing and checking results in the app uses none of your credits.
  • The credits are monthly. They renew on the 1st of the month (UTC). When they run out, further calls are refused until they renew, unless your workspace's owner buys top-up credits or upgrades. A per-minute rate limit, shared by the whole workspace, applies too.

Good habits for agent research

  • Ask for sources. Have the agent say which tool it called, with which filters, and list the domains or ids behind each finding so you can open them.
  • Keep it to the data. Tell it to use only numbers from tool results and to say plainly when a search returned nothing. That stops it filling gaps with guesses.
  • Verify what drives a decision. Before you order stock or brief a creative team, open the key stores in the app's Shops page and the key ads in its Ads page.
  • Remember what is estimated. Traffic, sales and ad spend for other brands are estimates. Use them to compare, not as exact figures.
  • Scope your prompts. One market, a result limit and a clear output use fewer calls than "tell me everything about pet products".
  • Read before you confirm. Tools that delete or share something only run with an explicit confirmation, so check what the agent proposes before you say yes.
  • Save prompts that work. A good teardown or ad research prompt becomes a template you reuse each week.

How to do this with SellWinners: agent, app and API

One API key covers all three ways in, and the agent is one of them. Use each for what it does best:

  • The web app for browsing and checking: filter stores on the Shops page, read ads on the Ads page, follow brands in your brand trackers and keep finds on your boards.
  • The agent over MCP for questions that need several searches stitched together: niche scans, teardowns, ad briefs and weekly summaries of what changed.
  • The REST API for jobs that run the same way every time, such as a daily export. The API docs list every endpoint.

To see the data behind each tool, start with Shops, Ads Library, Shop Analytics and Brand Tracker.

Key takeaways

  • MCP lets an AI agent call tools for live data instead of guessing from its training.
  • Connecting takes three values: the endpoint URL, the streamable HTTP transport and your API key in a header.
  • Specific prompts (niche, market, number of results, output format) get the best answers and use fewer calls.
  • A tool call costs the results it returns, from your monthly credits, shared with the REST API. Usage checks, lookups and failed calls are free.
  • Ask for sources, keep the agent to tool results, and verify key numbers in the web app before you act.