MCP for commerce engineers: connecting AI to your stack safely

Every commerce AI demo starts the same way: someone pastes an API key into a prompt and hopes. That is not architecture; it is a liability. The Model Context Protocol (MCP) offers a genuine answer, namely a standard way to expose tools, data, and actions to AI models with the same rigour we would apply to any API surface.

Why MCP instead of raw API access

Give a model your Admin API token and you have granted it everything the token can do, permanently, with no audit trail to distinguish the agent's actions from your own. An MCP server inverts this. You define exactly which operations exist, what their inputs look like, and what they return. The model sees capabilities, not credentials.

For commerce, that distinction is everything. "Read product catalogue" and "issue refund" should not sit behind the same permission.

Design tools around intents, not endpoints

The mistake I see in early commerce MCP servers is mirroring the platform API one to one, as a thin wrapper over every REST endpoint. Models work far better with intent-level tools such as get_slow_moving_inventory, draft_price_change, and summarise_customer_history. The result is fewer tools with richer semantics, each one mapping to something a merchandiser would genuinely ask for.

Writes need approval loops

My rule for production agents is simple: reads are free, and writes are drafts. An agent can read anything its tools expose, but every mutation, whether a price change, an inventory adjustment, or a customer refund, lands in an approval queue with the agent's reasoning attached. A human approves it. The agent receives feedback. Over time you loosen the loop for action types with a proven record, with spend and blast-radius limits enforced in the server rather than the prompt.

Start with the unglamorous wins

The highest-return agent work in commerce at present is not a chatbot on your storefront. It is operations: catalogue hygiene, anomaly detection on inventory and pricing, order exception triage, and report drafting. This work is repetitive, high in volume, and low in risk, which makes it ideal agent territory. Expose those capabilities over MCP, add an approval loop, and you will save genuine hours while your competitors are still debating chat widgets.