Guide to MCPs for Ecommerce Merchants

Anthropic introduced the Model Context Protocol (MCP) at the end of 2024. Just 2 years later, "Do you have an MCP?" has become one of the most common questions that ecommerce merchants ask of their software vendors.

MCPs have allowed ecommerce teams to replace their tedious and manual rituals of exporting reports from every platform they use and importing into AI assistants for analysis. Instead, they can query that data directly from their favorite AI application, be it ChatGPT, Claude, or something else.

To understand what separates a reliable MCP from a frustrating one, we talked to Daniel Villate, Sr Analytics Engineer at Postscript, who is part of the team that built the Postscript 1.0 MCP. Here’s what you should know about MCPs before you connect one and an introduction to using the Postscript 1.0 MCP.

MCP stands for Model Context Protocol, an open standard that lets AI assistants like Claude and ChatGPT securely connect to outside data sources. Think of it as a universal port: a platform builds one MCP server, and any compatible assistant can plug into it.

The Postscript 1.0 MCP is Postscript's connector built on that standard. Once connected, your AI assistant can pull your campaign, automation, acquisition, and shop-wide analytics and answer questions about them in plain language. For example, you can open a new chat in Claude and ask "Rank campaigns from the 1st to 15th of October by revenue and show sends, CTR, and unsub rate for each" and the agent will put together a report for you.

You can even connect multiple MCPs across your techstack (Postscript, Shopify, Meta, etc) and ask the AI assistant to query data from several platforms, such as “Show me campaign performance of my new product launch across every channel”.

There are two types of MCPs: read-only or read-and-write. Read-only MCPs can be given access to a platform and read the data within it to be able to answer your questions. Read-and-write MCPs allow the AI assistant to retrieve data as well as to take actions in your platform, such as creating, updating, modifying, or deleting. For example, with a read-and-write MCP, you could ask the AI agent to build a segment in your SMS platform, create a campaign around it, and send it.

But as more platforms rush to ship MCPs with as many tools built into them as possible, we keep hearing the same feedback from merchants who have tried them: even simple data reports can often be inaccurate so they don’t trust it to make any write actions. This could be due to user error (as with any prompt-based tool, there is some learning curve) but the way the MCP is built also has a significant impact on accuracy.

When a platform announces its MCP, the focus is often around a tool count. It’s an easy number to market, but it doesn’t tell you whether your assistant will get to the right answer.

A tool is one action your AI assistant can take in a platform, like pulling a campaign report. A skill is a set of instructions that tells the assistant which tools to use for a job and how to use them. “Tools are the routes an AI can take into your product,” Daniel explains. “A skill tells it which routes are right and when to use them. Lots of tools without that guidance leads to tool bloat.”

Tool bloat is what you feel as a user: the assistant tries several tools before it finds the right one, answers take longer, and every extra step uses more tokens, which can cost you more on your AI plan.

Tool counts are also easy to inflate. “It’s easy to break things into more tools,” Daniel says. “A tool that just looks up an ID isn’t valuable on its own. It only matters combined with others.” When you compare MCPs, ask how reliably each one gets to the right answer, not how many tools it has.

Before you connect any MCP, there are a few questions worth asking:

• Who can see your data? Look for a secure sign-in that ties the connector to your account only. Postscript 1.0 MCP: You sign in to Postscript and approve access, so your agent only sees the shop tied to your sign-in.

• Where does your data go? Ask whether your data is used for training and what the vendor logs. Postscript 1.0 MCP: We don’t use MCP data to train AI models, and we never see your AI conversations. Request logs stay in Postscript’s systems and never include subscriber contact info or credentials. Once results reach your AI assistant, that provider’s data settings apply, so review them before connecting.

• Can you trust the numbers? Some “wrong number” moments come from definitions, like attribution windows or what “last month” means. Others come from how the MCP is built: whether your assistant can find the right data, and whether the data is clear once it gets there. Ask the vendor how they tested for accuracy. Postscript 1.0 MCP: Queries run directly against your Postscript data, not estimates, and we document how each metric is calculated, so any difference from the app has an explanation. We tested how reliably assistants reach the right data before we tested anything else (more on that below). If something looks off, tell your agent “give feedback” and it goes straight to our team.

• Can it change anything in my account? Reading and writing are very different levels of risk. A wrong report is an annoyance, but a wrong action can send a campaign or edit a live flow. Postscript 1.0 MCP: It’s read-only. It can’t create, edit, send, or delete anything in your account.

• What are the limits of the MCP? Knowing what the MCP should not be used for. For example, an AI agent can give you inaccurate strategy advice without knowing the full context of how your connected platform works and your goals. Postscript 1.0 MCP: Use it to pull the numbers, then bring “should I change this?” questions to your Postscript team for expert advice. As Daniel puts it: “It’s easy to assume an MCP is plug and play: you connect it, ask questions, and get answers. But marketing platforms are nuanced. Expecting an AI assistant to know how everything in your platform works and have all the right opinions is misleading.”

Our goal with the Postscript 1.0 MCP was to give merchants SMS numbers they can trust. That goal shaped the most important decision: limiting the scope to read-only. A read-and-write MCP gets complicated fast, and that complexity shows up as mistakes in your account.

Getting there took three versions. Each time, Daniel tested one thing before anything else: could the assistant reliably find the right data? “I tested routing accuracy first, before data accuracy,” he says. “If the agent can’t find the right place to look, it doesn’t matter how accurate the data is.”

The first version had a separate tool for every report. “Once we hit 15 to 20 tools, the agent spent more time figuring out which one to call,” Daniel says. “That meant slower answers and higher token costs for merchants.” The version we ultimately launched replaces those single-purpose tools with a few flexible ones. Instead of one tool for campaign reports, another for automation reports, and another for popup reports, there’s one report tool. Your assistant tells it what to pull, like campaigns from June 1 to June 30, and the tool returns that report. With fewer tools to choose from, the assistant takes fewer wrong turns and gets to the right data faster.

Once the assistant could reliably find the data, the next step was making sure it couldn’t misread it. “Every metric and every dimension is unique, so there’s nothing for the AI to mix up,” Daniel says. “Where a number needs context, like how attribution is defined, we pass that note along with the data so your assistant shares it with you.”

After both internal testing and customer feedback, we’re very confident in the accuracy of the Postscript 1.0 MCP.

The Postscript 1.0 MCP connects your SMS analytics to Claude (web, desktop, and Claude Code), ChatGPT, and Codex. It's available to all Postscript customers. 

Here's what your agent can pull:

• Campaign, automation, and popup performance. Sends, clicks, CTR, conversion rate, orders, revenue, and unsubscribes for each. Try: "How did my popups perform from July 1 to July 31?"

• Shop-wide trends. Attributed revenue, messages sent, and new subscribers by hour, day, week, month, or year. Try: "Show my new subscribers by month for the last 12 months."

• Rankings. Top or bottom performers by any measure. Try: "Sort my campaigns by revenue per send from June 1 to June 30 and show the bottom five."

• Exports. Full tables you can drop into a spreadsheet. Try: "Give me every campaign and automation from June 1 to August 31 in one table: IDs, sends, clicks, orders, and revenue."

Let’s say you need a weekly recap. Instead of exporting reports or screenshotting dashboards, you ask your agent: "Compare sends, revenue, and CTR for Oct 12–18 vs. Oct 5–11." It pulls the numbers from your Postscript account and lays them out side by side, next to the Shopify and Meta data you've already connected.

Learn how to connect to the Postscript 1.0 MCP here

The best questions name a metric, a date range, and a scope. Get those three right and you will rarely need a follow-up.

• Name the metric. Say "revenue," "sends," "CTR," "unsub rate," or "orders" rather than "how did we do." If you want several, list them: "sends, revenue, and CTR."

• Use explicit dates. Say "June 1 to August 18" rather than "the last 90 days." Explicit dates return the most reliable results.

• Say what to compare. "Sep 1–30 vs. Aug 1–31" gets you a side-by-side instead of one number.

• Say shop-wide or by entity. "Total SMS revenue" pulls one shop-wide number. "Revenue by campaign" or "which automation drove the most" pulls a per-entity report. These use different revenue definitions (see "Know the numbers" below).

• Use the names you know, loosely. Say "the welcome flow" or "the Labor Day campaign." The assistant looks up the exact name first, so you don't need to match it.

• Say how you want it delivered. "Give me a 3-bullet summary for my exec update" or "put it in a table" saves a round of edits.

• Ask it to verify. Good follow-ups are "Is that the full result, or was anything cut off?" and "Do those rows add up to the total?"

Weak: "How's SMS doing?"

Strong: "Compare SMS revenue, sends, and CTR for Sep 1–30 vs. Aug 1–31, and name the top 3 campaigns by revenue."

Prompts to steal, by role:

• Retention: "Which campaigns had high CTR but low conversion rate from [start] to [end]?"

• Growth: "Rank my popups by conversion rate from [start] to [end]."

• Executives: "Summarize [last quarter] vs. [prior quarter] in 3 bullets: revenue, sends, and top campaign."

• Agencies: "Build a monthly SMS report for [month] vs. [prior month]: revenue, sends, CTR, top 5 campaigns, top 3 automations, and new subscribers."

We recently announced Postscript 2.0, a rebuild of our entire platform to be AI-native. Every line of code on this new foundation is built so both humans and AI agents can work with it, which means the Postscript 2.0 MCP isn’t limited by 10-year-old tech. Instead of handing an assistant every tool at once, the 2.0 MCP organizes them by area, so the assistant goes step by step: first to the area it needs, like campaigns, then to the tools for that area. That’s how it handles complex read and write actions without the tool bloat we covered earlier. For example, one customer connected the 2.0 MCP to Codex, then had Codex build a form, create a segment around it, and write an email automation it triggers. We’ve already started migrating customers to Postscript 2.0, and those customers have access to the 2.0 MCP.