chatgpt ecommerce integration, claude connectors, ai for lumber yards, mcp, building supply ecommerce
ChatGPT ecommerce integration for dealers needs an undo button
October 30, 2026
How a lumber or building supply dealer connects ChatGPT or Claude to a ToolSwift store, what it can read and change, and six guardrails to check first.
Executive summary
A ChatGPT ecommerce integration for a ToolSwift dealer means adding the store's own connector URL to ChatGPT or Claude and approving it on a consent page. The assistant then answers questions about sales, orders, quotes and customers, and proposes edits you preview and approve, undoable for 30 days. Fields the ERP owns are refused.
Most pages about ChatGPT and eCommerce show you how to plug an assistant into a store. None of them shows you how to undo it.
Most of what ranks for this phrase is about shoppers buying inside ChatGPT. This page is the other half: your staff working your own store through ChatGPT or Claude.
I run ToolSwift, which sells eCommerce to lumber yards, hardware stores and building supply dealers, and we have just put a Claude and ChatGPT connector into dealers' hands. So I have a stake in this. I also have a view on what that connection should be allowed to touch, and it is narrower than most guides assume.
A dealer's store sits on top of an ERP that owns the prices and the stock (and usually the credit limits). An assistant that writes into those fields is writing into the wrong system. Everything below follows from that.
What can ChatGPT or Claude do in a ToolSwift store?
It answers questions about your store, and it proposes bulk edits that you preview and approve before anything is written. Every approved change is logged and can be undone for 30 days, and fields the ERP owns are refused.
Once connected, the assistant can answer questions about sales, orders, quotes, customers, credit and categories. "Which contractors have quotes open over a week?" or "Which categories sold less this month than last?" are read questions; nothing in the store changes when you ask them.
When you ask for a change, the assistant proposes it and you preview and approve it before anything is written. It can propose bulk edits to products, customers, companies, store policies, shipping settings, category content and image alt text. Price edits are off by default; an owner can switch them on, and each price change then needs approval. Every approved change lands in the activity log, and can be undone for 30 days; records changed again since are skipped. If the request touches a field the ERP owns, the connector refuses it.
In my view, that list covers the jobs a yard tends to put off: alt text on a few thousand product photos, category pages with no copy, company records with three spellings of the same contractor, store policies nobody has reread since launch. These are content and record jobs, and none of them needs to move a price.
Two more pieces are coming, with no date yet: reports and exports through the assistant, and AI quote drafts that only your staff can send. The second one matters to me. An assistant may prepare a quote; a person at the yard decides whether it goes out.
The connector sits next to the Takeoff Agent, which is live and turns a contractor's voice note, photo, scan, takeoff or file into a draft quote for staff to review. The rest of what AI does inside the platform is on the AI features page. Assist AI, the separate assistant for staff email, is still in review and is not part of this connector.
How do you turn on the connector in ToolSwift?
Five steps: switch connections on, copy your store's connector URL, add it to Claude or ChatGPT, click Connect, then Allow on ToolSwift's consent page. A dealer connects their own assistant to one store at a time.
- In the ToolSwift dashboard, open Settings, then Connected apps, and turn on "Allow assistant connections".
- Copy the connector URL shown on that screen. Each store has its own URL, and the URL fixes which store the assistant reaches. The same screen has the steps for Claude and for ChatGPT.
- In Claude, add the URL as a custom connector. In ChatGPT, add it as a custom MCP server. The next section has the clicks for each.
- Click Connect. ToolSwift's consent page opens, on a phone as well as a computer. It shows which assistant is connecting, that it connects as you, the store, what the assistant can do, and what it can never do: delete, take payments or refunds, message customers, change staff or roles, or see cost, margins or cards. Prices are off unless the store owner switches them on, and then every price change waits for approval.
- Click Allow, or Deny if anything on that page surprises you.
Connected apps then lists each connection with the app, the person, the access, when it was connected and when it was last used, and a Disconnect button on every row.
Turning off "Allow assistant connections" stops every existing connection at once. That is the switch I would want within reach the day someone leaves.
How do you add the connector in Claude and in ChatGPT?
In Claude it is a custom connector; in ChatGPT it is a custom MCP server. Both assistants speak the Model Context Protocol (MCP), the open standard Anthropic open-sourced on November 25, 2024 for connecting AI apps to outside systems.
In Claude, go to Customize, then Connectors, click "+ Add", choose "Add custom connector", and paste your store's connector URL with a name you will recognize. Anthropic's custom connector guide says custom connectors work on Free, Pro, Max, Team and Enterprise plans, and that "Free users are limited to one custom connector." On Team, only Owners and Primary Owners can add one; members then connect from Customize, then Connectors.
To use it in a chat, click the "+" button (or type "/"), hover over Connectors and switch the connector on for that conversation. Anthropic's connector build docs cap a tool result at about 150,000 characters and a tool call at 240 seconds on claude.ai and Desktop, so ask about one category or one month at a time rather than the whole store.
In ChatGPT, OpenAI's connect guide says to open Plugins, press the plus button, choose "Add custom MCP server", and enter a name, a description and the connection, which is your store's connector URL. You set up sign-in, accept a risk warning and review the tools ChatGPT found. On which accounts can do this, the page says only: "Account and workspace policies apply to adding and using custom MCP servers."
What should you check before any assistant writes to your store?
Six things: a consent step per store, a preview of the exact change, approval by a person, a log with undo, a hard refusal on ERP-owned fields, and one place to revoke access. If a vendor cannot show all six on your own catalog, keep the connection read only.
The assistant makers' own safeguards stop at the chat window. OpenAI's MCP guide says "ChatGPT currently requires manual confirmation in any conversation before write actions can be taken," and notes that write actions can still happen when a server labels a tool as read only. It also says custom MCP servers "are not developed or verified by OpenAI."
Anthropic goes further. On Team and Enterprise, Owners can set each connector action to Always allow, Needs approval or Blocked across the company, and users cannot override it, according to Anthropic's connectors guide. The same guide says "Claude inherits each person's permissions from the connected service," so a sales rep's Claude can do what that rep can do in the store. Anthropic's custom connector guide warns that malicious servers can hide instructions, and tells you to switch off tools that write to outside apps when you use Research, because Claude can call them there without asking again.
All of that protects you inside the chat. None of it helps after you click yes.
A confirm button tells you the assistant is about to write. An undo button is what you need when it was wrong.
The ERP is the second reason. Your ERP (Spruce, BisTrack, Agility, Paladin or another) owns the price, the stock, the customer's credit limit and the account number. The website shows them. If an assistant writes a new price on the website, one of two things happens: the next ERP sync overwrites it, or the website and the ERP disagree until a contractor gets charged the wrong number at the counter. Either way, the assistant changed the wrong system.
The two tests that separate vendors are undo and the ERP refusal. A preview is the easy part to build. Undo means the tool kept the old value of every field it touched, and the ERP refusal means it knows which of your fields came from the ERP in the first place.
Run the ERP test on purpose. With price edits switched off, ask the assistant to drop the price of a 2x4 or raise a contractor's credit limit. The right answer is a refusal. A wrong answer is a cheerful "done."
General platforms take a different route. Shopify's connector for Claude writes catalog changes, variant prices included, once you grant write access, keeps orders and customers read only, and "changes items one at a time, so a large batch can stop partway," leaving the rest to finish in the admin.
What should you let the assistant do in week one?
Ask questions in week one, let it write content in week two, and let it touch customer records only after you have reversed at least one of its changes yourself.
The order matters because each step raises the cost of a mistake. This is the sequence I would give any dealer:
- Week one, read only. Ask about sales, open quotes, customers with credit questions, slow categories. Check two or three answers against your own reports. If the numbers do not match, stop here and find out why.
- Week two, low-risk content. Image alt text and category descriptions. A bad alt text costs you nothing at the counter. Approve a small batch, read the preview line by line, then undo one change to prove you can.
- Week three: store policies and shipping settings. These are short, and a customer reads them. Have the person who wrote them approve the edit.
- After that: customer and company records. Merging duplicate contractor records is valuable and easy to get wrong. Do it in batches small enough to check by eye.
- Leave prices to the ERP. If your ERP sets prices, keep "Allow price edits through assistants" switched off, so price, stock, credit limits and account numbers change in the ERP and flow out to the website.
Two hygiene rules from Anthropic's own guide apply here too. Connect only servers built by companies you trust, and on a Team or Enterprise plan, a connector with one shared key gives every user that key's access, so limit the key to what members should reach. On ToolSwift, disconnect that person's row in Connected apps the day they leave.
Where does this approach cost you?
Approval is slower than a direct write. If you want 500 products changed in one sentence with no review, a preview-and-approve connector will feel like a clerk asking you to initial every page.
That is the trade, and I would make it every time, but it is a real cost. There are three others worth knowing before you start.
Price edits through the assistant are off by default. An owner can turn on "Allow price edits through assistants" (it needs the Pricing permission), and even then every price change waits in Settings > Connected apps > Activity for someone to approve it, before and after shown side by side. For a dealer whose ERP sets the prices, the right setting is off, which means a bulk reprice still happens in the ERP and not in a chat window.
Today the assistant reads, answers and proposes edits; reports and exports through the connector, and quote drafts, are coming with no date.
And a preview is only as good as the person reading it. A tired manager approving a 300-line preview at 6 p.m. is a confirm button with extra steps. Keep batches small enough that someone actually reads them.
My rule for any dealer: give an assistant write access to your store on the day you can watch it undo a change, and never give it write access to anything your ERP owns. If you want the broader argument for why dealers will pick software on trust rather than AI features, I made it in Will AI commoditize software?.
If you want to watch the connector preview an edit, wait for your approval and then undo it on a catalog like yours, book a ToolSwift demo.
FAQ
Is it safe to connect ChatGPT or Claude to my store?
It is safe when a person previews and approves every change, when changes can be undone, and when fields your ERP owns are refused. In ToolSwift, each change is logged and can be undone for 30 days. Both OpenAI and Anthropic warn that custom connectors are not verified by them, so connect only servers from companies you trust.
Can the assistant change my prices?
Only if the store owner allows it. "Allow price edits through assistants" is off by default, turning it on needs the Pricing permission, and every price change still waits for your approval. If your ERP sets your prices, leave it off and reprice in the ERP so the website follows.
What if the assistant gets something wrong?
Undo it. Every approved change lands in the activity log and can be undone for 30 days. If a record was changed again after the assistant's edit, the undo skips that record so it does not overwrite the newer change.
Who can connect an assistant to a ToolSwift store?
Store staff who administer that store. The consent page only offers stores the signed-in person administers, the assistant connects as the person who clicked Allow, and Connected apps shows each connection with that person's name, so the owner can see who connected what and disconnect any row.
What are Claude connectors?
Claude connectors link Claude to an outside service through the Model Context Protocol, so Claude can read data and take actions there. You add one from Claude's directory, or by URL as a custom connector. Claude inherits each person's permissions from the connected service.
Sources
Related in ToolSwift
Connected to the systems you already run
Sixteen ERP and accounting connections, and our team does the setup.