Classify products from Claude and ChatGPT

Questions that used to go to a spreadsheet or a broker now often go to an AI assistant first: “What’s the HS code for this?”, “Will my customer pay duty on that?” Assistants understand the question well. On their own, they are not a reliable source of tariff codes: they may recall outdated schedules, mix up countries or invent a plausible-looking number.

The border.bot MCP server gives the assistant a tool to call for those answers, so the code comes from a classification engine instead of the model’s memory.

What MCP is#

The Model Context Protocol (opens in a new tab) is an open standard for connecting AI assistants to external tools and data. An MCP server describes the tools it offers; the assistant decides when to call them, sends structured inputs and gets structured results back.

A remote MCP server runs on the web. You add it to your client with a URL and sign in through your browser with OAuth. Nothing is installed locally and no API key goes into a config file.

What the border.bot server can do#

Once connected, your assistant can:

• Classify products from a description or a product URL, for any destination, returning the code in that country’s format with official descriptions, a confidence score and alternatives.
• Read product pages to pull out the title, materials, price and country of origin before classifying.
• Calculate landed cost for a code, origin and destination, with each duty, tax and fee and the total.

Every call runs on the same engine as the border.bot dashboard and API, uses your workspace’s credits at the same prices, and is saved to your workspace history.

Setting it up#

The server URL is https://api.border.bot/mcp. In short:

• Claude: Customize › Connectors › “+ Add” › “Add custom connector”, enter the URL, then sign in to border.bot.
• ChatGPT: on chatgpt.com/plugins, choose the plus button › “Add custom MCP server”, enter the URL with OAuth, then install the plugin and sign in.
• Codex: run codex mcp add borderbot --url https://api.border.bot/mcp, then codex mcp login borderbot.
• Cursor: add the URL to ~/.cursor/mcp.json (or use “Add to Cursor” on the MCP setup page), then complete the sign-in Cursor asks for.

When you sign in, you approve the connection and choose which workspace pays for the calls. Menu names change between client versions, so check the MCP setup page for current step-by-step instructions, including Claude Code, Visual Studio Code, Gemini CLI and the other clients we have guides for.

What to ask#

Some prompts that work well:

What’s the US HTS code for a men’s 100% cotton crew-neck T-shirt made in Vietnam? Show me the description at each level.

Classify this for import into Germany and tell me where it’s made: https://shop.example/products/merino-beanie (opens in a new tab)

Calculate the landed cost of 50 units at $18 each, HS 4202.92, from China to the UK with $120 shipping.

Here are twelve product descriptions from our catalogue. Classify each one for Canada and list any that come back with low confidence.

The last one shows where an assistant shines: it can loop over a list, call the tool for each item and summarise the results, while the codes themselves still come from the tool.

What happens behind a single question#

Take the first prompt above. The assistant recognises that it needs a classification, so it calls the border.bot classify tool with structured inputs: the description, the destination (US) and the origin (VN). border.bot returns a structured result with the 10-digit HTS code, the official description for the chapter, heading, subheading and tariff line, a confidence level, the reasoning and the closest alternatives. Your workspace is charged for one classification.

The assistant then writes its answer from that result. Because the code, the descriptions and the confidence come straight from the tool, you can check the assistant’s summary against them, and the same result is waiting in your workspace history if you want to share it with a colleague or re-run it later.

Tips for reliable answers#

• Ask it to use border.bot. Say “use the border.bot tool” so the assistant doesn’t answer from memory.
• Give it the facts customs needs. Material, function, construction and who the product is for matter more than the product name.
• Always name the destination. Codes beyond six digits differ by country.
• Ask for alternatives on low confidence. If the result is flagged for review, ask the assistant to show the alternatives and what would distinguish them, then add that detail and run it again.
• Keep origin honest. If the assistant reports an origin detected from a product page, confirm it with your supplier before you rely on it.

Billing and security#

• You sign in on border.bot; the assistant receives a scoped OAuth token, never your password.
• Each connection is tied to the workspace you chose, and its calls draw on that workspace’s credits. If a call fails, its credits are refunded automatically.
• You can see every connected client in the dashboard and revoke any of them at any time.

Where this fits#

The MCP server suits ad-hoc questions, research and one-off batches inside the tools your team already uses. For checkout and catalogue automation, use the REST API; for reviewing and sharing results, use the dashboard. All three use the same engine and the same credits.

To see the results before connecting anything, try the free HS code classifier. It runs on the same engine.