What BigCommerce changed on November 19
BigCommerce has opened Store Management MCP to every store. Until this week, a merchant who wanted an AI agent to work inside the store admin had to fill in the Commerce AI Labs form and wait for approval. The November 19 changelog says that step is gone: any store can now turn it on, and the product stays in early access.
MCP (Model Context Protocol) is the open standard AI clients use to call outside tools. Store Management MCP is the merchant-side server. It lets an agent such as Claude, Cursor or ChatGPT read and change catalog, pricing, inventory, orders, promotions and store profile data. Storefront MCP, which serves shoppers, is a separate product, and BigCommerce reorganized its MCP overview the same day to cover Storefront MCP, WebMCP and Store Management MCP side by side.
Two practical changes shipped with the wider access. A single server address, https://api.bigcommerce.com/mcp, now works for all stores, with a store picker after sign-in. Agencies and multi-store merchants can also connect straight to one store at https://api.bigcommerce.com/stores/{store_hash}/mcp. Rate limits are now set per user and per store, and an over-limit response carries a Retry-After header so the agent knows how long to wait.
What an agent can do through the 16 tools
The Store Management MCP reference lists 16 tools. Product tools cover lookup, search, creation, category assignment, content edits, custom fields, catalog prices and price list records. Inventory tools read and set stock levels. Order tools list orders, quote a refund and create a refund. The rest list and create promotions and read store information.
Each tool has hard per-call limits. A single update_catalog_prices call handles up to 10 products and 50 variants. update_price_list_records takes up to 1,000 records, and set_inventory_levels takes up to 2,000 items. When an agent asks get_orders for line items, the page size drops from 250 to 25. Those numbers matter if you picture an agent repricing a whole catalog from one prompt. It will work through the job in batches, and each batch counts against the user’s rate allowance.
BigCommerce has built in some caution. Products an agent creates are hidden, and promotions it creates start disabled, so a person still publishes them. Coupon codes cannot be created or listed through the tools at all.
How access and scopes work
Only the store owner can switch the server on, under Settings, then API and MCP, then MCP, by accepting the beta terms. The store must be active. Connections use OAuth 2.1 with PKCE. Clients register themselves, so there is no client ID or secret to manage, and the agent never holds a password or API token.
Access is granted per user, per store. Every new user starts at None in every scope, and an administrator raises them one area at a time.
- Products, Inventory, Promotions and Orders each have None, read-only and modify levels.
- Store is None or read-only and exposes business contact details.
- Customers is None or read-only, and no tool uses it yet.
- Orders modify allows refunds.
Owners can hand scope management to other staff by granting the View and Create AI MCP Scopes Grants permissions. Turning the server on or off stays with the owner. The reference lists Claude Desktop, Claude in the browser, Claude Code, Cursor, ChatGPT and Grok as tested clients, with plan requirements for the last two.
Limits worth knowing before you connect anything
Refunds are the obvious risk. Orders modify lets an agent call create_refund, which moves real money and cannot be reversed through the tool. BigCommerce never retries that call automatically, which helps, but a wrong order number in a prompt is still a wrong refund.
Scopes cover a whole area, with no field-level control. Every agent a user connects shares that user’s access, and activity is logged against the user with no separate agent attribution. Scope changes can take up to the remaining session length to apply, and the reference says there is no self-service view of live connections and no way to force a disconnect. If you remove someone’s access in a hurry, assume their agent can keep working until the session ends.
What we would set up first
Our advice is to start read-only. Create a dedicated staff user for agent work, give it read access to Products and Inventory, and run it against a sandbox store if you have one. Ask it the questions your team answers by hand today: which variants are below a stock threshold, which products are missing descriptions, which price list records disagree with the catalog price.
Move to modify on Products or Inventory only after someone has checked a week of its answers. We would keep Orders at read-only for most teams and leave refunds with people, because the tool cannot undo them and the logs will not tell you which agent made the call. Write down who holds which scope, since BigCommerce does not yet give you a list of live connections to check against.
If you build apps or internal tools on top of it, handle the Retry-After header from the start. All agents under one user and both server addresses share the same allowance, so a scheduled job and an interactive chat run by the same person will compete for it.





