Ask your AI agent about your store, and approve every change it makes
Benson's MCP server brings your extension numbers into Claude, ChatGPT, Cursor and other agents. Every change shows a preview and waits for your yes.
Jump to a section
In short
It's Sunday evening and you're planning next week in Claude. The sale calendar is there, the email drafts, the stock you need to clear. Then a question comes up that matters: did coupon extensions eat into the margin on the last sale?
Claude can't tell you. That number lives in another dashboard, one you open once a month. So you can open a new tab, sign in, find the right page, export a CSV and paste it back. Or you can carry on planning and leave the question for later. Most weeks, later never comes.
Benson now has an MCP server. Connect it to the AI agent you already use, and that question gets an answer from your own store's data, in the conversation you're already having. Your agent can also make the everyday changes in Benson, but only after it shows you a preview and you say yes.
Why do the useful questions go unasked?#
Because the answers are spread across too many places. Shopify said back in 2021 that the average merchant had six apps installed on their store [Shopify Partners, Feb 2021] (opens the source), and each one tends to bring its own dashboard and login. Add the ad platforms, email, analytics and the accounts package, and checking one number is a small project.
72%
of marketers in Funnel's 2026 survey said they have mountains of data but find it hard to turn into insight.
Funnel sells marketing data tools, so it has an interest in that answer, but merchants describe the same thing in their own words.
Meanwhile, the place where owners do their thinking has moved.
58%
of US small businesses said they use generative AI in 2025, up from 40% the year before.
So the plan gets made in one window and the numbers sit in another. Every question that needs both costs a trip, and the questions that only matter a little never get asked.
Why is this happening now?#
Because AI tools finally share a way to reach other software. Anthropic released the Model Context Protocol (MCP) as an open standard in November 2024. Its announcement named the problem plainly: every new data source needed its own custom integration [Anthropic, Nov 2024] (opens the source).
The rest of the industry followed. By December 2025 Anthropic had handed MCP to the Agentic AI Foundation, a new Linux Foundation body launched with projects from OpenAI and Block too, and with Google, Microsoft and AWS among its members. The foundation said ChatGPT, Gemini, Microsoft Copilot, Cursor and VS Code had all adopted it [Linux Foundation, Dec 2025] (opens the source). Google announced official MCP servers for its own cloud services the same month [Google Cloud, Dec 2025] (opens the source).
10,000+
published MCP servers a year after launch, according to the Linux Foundation.
For a store owner, that means one connection standard works across the assistants you're most likely to use. For an agency, it means the same agent can work across client accounts, which is useful and also where mistakes creep in.
Why would you let an AI change anything in a live store?#
That's the right question to ask, and most businesses are asking it.
27%
of executives at large companies trusted fully autonomous AI agents in 2025, down from 43% a year earlier, in Capgemini's survey.
That survey covered companies with more than $1 billion in revenue, not small stores, but the worry travels. An agent can misread a request. It can act on the wrong account. And text it reads, from a web page or a campaign name, can try to steer it.
The security guidance is consistent about what to do. OWASP's list of AI risks says to give agents only the permissions they need and to make a human approve high-impact actions before they happen [OWASP GenAI Security Project, Nov 2024] (opens the source). The MCP specification itself says there should always be a human in the loop who can deny a tool call [Model Context Protocol, Jun 2025] (opens the source). Anthropic and OpenAI both tell users to connect only servers they trust, because a malicious one can try to inject instructions [Claude Help Center, Oct 2026] (opens the source) [OpenAI Help Center, Oct 2026] (opens the source).
So the answer isn't "never". It's narrow permissions, a preview before every change, and a person who says yes.
What can you do about it today?#
There are three honest routes, and each has a catch.
Export and paste. Most dashboards, Benson's included, export a CSV. Download it, drop it into the chat and ask away. It costs nothing and needs no setup. The catch: it's a snapshot that's stale by tomorrow, it's manual every time, the file may carry customer data into a chat you didn't plan to share it with, and your AI can read it but can't act on it.
Build your own connector. If you have a developer, or patience, you can write a small MCP server against your store platform's API. The catch is the work around it: credentials, scopes, keeping up when the platform changes, and deciding what your agent may change. And your platform's API holds only what the platform records, like orders and the codes used on them. It doesn't know which extensions were in a shopper's session.
Use the vendor's MCP server. This is the least work, and it means trusting the vendor as well as your agent. Before you connect one, check five things. Can you connect read-only? Is each connection tied to one account, so it can't drift onto another client's data? Is approval for changes enforced by the server, not just requested by the chat app? Can you see each connection's calls and revoke it? And is text from shoppers treated as data, not instructions?
One more catch applies to every route. A connection is plumbing, not judgement.
That's fair. MCP gets the numbers to your agent. Whether the answer is useful depends on the data behind it and on how carefully you read it.
What Benson's MCP server does#
Add one URL to your agent and sign in to Benson. There are no keys to copy. On the consent page you pick one organisation and choose what the connection may do: read only, or read and make changes after you confirm each one.
| Scope | On the consent page | Covers |
|---|---|---|
benson:read | Read your store's data | Stores, install status, analytics, extensions, orders, insights, rules and team. |
benson:write | Make changes, after you confirm each one | Set up stores, change the pop-up blocker and checkout rules, edit and publish offers, and invite teammates. |
It works with any MCP client that supports remote servers and standard sign-in. These have step-by-step setup:
Set up. Give your agent your store's address and it can add the store, get the install snippet and the steps for your platform, check the storefront to confirm Benson is live, and apply the recommended setup. It can also start a short test session that shows Benson catching a simulated coupon pop-up on your own store.
Read. It can answer from most of what the dashboard shows: which extensions are in your sessions and what they cost, your discount numbers, orders and the codes used on them, what Benson has protected, measured against the holdout, AI shopping agents on the store, affiliate overwrites, live activity, insights, your settings, offers, experiments, your team, playbooks and guides. Reads change nothing.
Change. It can switch the pop-up blocker between observe and block, start or stop a holdout, add and edit rules, change your checkout rules, draft and publish offers, act on insights and invite teammates. Every one of these asks first.
Three ready-made prompts give you a starting point:
| Prompt | In Claude Code | What it does |
|---|---|---|
set_up_store | /mcp__benson__set_up_store | Add the store, install Benson, check it's live and apply the recommended setup. |
weekly_review | /mcp__benson__weekly_review | What changed this week: extension traffic, its cost, new insights and what to do next. |
stop_coupon_popups | /mcp__benson__stop_coupon_popups | Measure first, then block the extensions that cost the most, with a holdout to prove it. |
How a change works#
When your agent asks Benson to change something, the first answer changes nothing. It's a preview: what would change, on which store, from what to what, and who it affects. Your agent shows you the preview, and only goes ahead when you say yes.
That yes travels as a single-use token. It works once, for that exact change on that connection, within ten minutes. Different arguments, an expired token or a token from another connection all fail. If a retry arrives, Benson returns the first result instead of acting twice.
It can't do more than you can. Every call goes through the same checks as the dashboard: your role, your plan and your organisation. Members can only read. If a change needs a plan you don't have, the preview says so before you're asked anything.
Some things never go through an agent. These stay in the dashboard, behind its own confirmations:
- Billing: choosing a plan, checkout, cancelling
- Deleting a store, purging its data, or deleting the organization
- Rotating site keys, policy keys or edge secrets
- Removing members, changing roles or transferring ownership
- Connecting integrations that need a secret (email platforms, Slack)
- Exporting all of the organization's data
You can see it and stop it. Settings → AI agents lists every connection, what it can do, when it was last used and its recent tool calls. Disconnect stops it on its next call. Changes made through an agent are recorded as made via that client, next to the name of the person who said yes.
Shopper text stays data. Some of what Benson reports was typed by someone else: UTM tags, campaign names, discount codes. Benson marks it as untrusted and tells your agent to treat it as data, not instructions. That lowers the risk of a code or a page steering your agent. It can't remove it, which is one more reason changes wait for you.
For agencies and scripts#
Each connection works in one organisation, and no tool takes an organisation as an argument, so a connection can't wander into another client's account. Agencies connect once for each client. For agents that run without a person, like scheduled jobs, you can create an agent key instead. It's shown once, works in one organisation and can be revoked any time. A headless agent still gets previews, so decide in your own code, or with a person, before you confirm.
Plans#
The MCP server works on every Benson plan, Free included, and your agent can do what your plan includes. Your AI tool has its own rules too. Claude's free plan allows one custom connector [Claude Help Center, Oct 2026] (opens the source). OpenAI currently offers custom MCP apps that make changes on ChatGPT Business, Enterprise and Edu, and read-only connections in developer mode on Pro [OpenAI Help Center, Oct 2026] (opens the source). Check your plan with your AI provider before you rely on the change tools.
We're preparing listings in Claude's connector directory, ChatGPT's apps and the official MCP Registry, and we'll add tools as the dashboard grows. Each new one follows the same rules: reads change nothing, and changes wait for your yes. If there's something you'd like your agent to do in Benson that it can't yet, tell us at [email protected].
The MCP server page has the setup for each client, and the MCP server reference covers every tool, the confirmation protocol, errors and limits. A good first prompt is the weekly review: ask your agent what changed this week and what you should do about it. Not using Benson yet? Install the Shopify app if you're on Shopify, or see Benson on your store for WooCommerce, BigCommerce, Magento and custom stores.
Questions merchants ask
- Can I connect it read-only?
- Yes. Turn off the option to make changes when you connect, and your agent sees only the reading tools. Members of an organisation can only connect read-only.
- Can my agent change things without asking me?
- No. Every change answers first with a preview and nothing happens until you say yes. The one exception is starting a short test session, which creates a test link and changes nothing in your store.
- Can it touch billing or delete my data?
- No. Billing, deleting a store or its data, rotating keys, removing members, changing roles and exporting all your data stay in the dashboard. It can delete a single blocking rule or cancel an invitation, after you confirm.
- I run an agency. How do I connect several clients?
- Connect once for each client. Each connection is bound to the one organisation you pick when you sign in, so your agent can't mix up which account it is working in.
- Is my data used to train models?
- Benson doesn't train models on your data, and it only sends data to the agent you connect, when that agent asks. What your AI provider does with it is set by their terms and your settings with them.
Sources#
- Anthropic: Introducing the Model Context Protocol (Nov 2024)
- Linux Foundation: Formation of the Agentic AI Foundation, anchored by MCP, goose and AGENTS.md (Dec 2025)
- Google Cloud: Announcing official MCP support for Google services (Dec 2025)
- Model Context Protocol specification: Tools (2025-06-18)
- OWASP: LLM06:2025 Excessive Agency
- Capgemini Research Institute: Trust and human-AI collaboration set to define the next era of agentic AI (Jul 2025)
- U.S. Chamber of Commerce: Empowering Small Business report, 2025 (Aug 2025)
- Funnel: The 2026 Marketing Intelligence Report
- Shopify Partners: How to build a Shopify app (Feb 2021)
- Claude Help Center: Get started with custom connectors using remote MCP
- OpenAI Help Center: Developer mode and MCP apps in ChatGPT
- Benson: MCP server
- Benson docs: MCP server reference
- Benson: How to use a holdout to measure coupon controls



