TL;DR
- Ask the model for a research brief and candidate list, not a final or "broker-grade" classification.
- Connect an authenticated, read-only MCP catalog so the model can search current HTS rows, verify that proposed codes exist, and retrieve hierarchy and notes.
- Make the model state missing product facts and unresolved conflicts. Never let it invent composition, use, origin, dimensions, or other facts needed to choose a heading.
- Have a qualified reviewer apply the legal sequence and approve the code. Run duty and AD/CVD as separate consequence checks.
Set the boundary before you write the prompt
A polished answer can still be wrong. A model may invent a statistical suffix, skip an exclusion, treat a search result as a legal conclusion, or reason from product facts that were never supplied. Tool access reduces some of those failure modes, but it does not eliminate classification judgment.
Tandom's public MCP surface deliberately fails closed to a small read-only catalog. It includes HTS search, exact lookup, hierarchy, notes, duty, PGA, Chapter 99, AD/CVD, and regulatory-data status. It does not expose the private classifier, document extraction, broker overrides, or write actions.
The right deliverable is a traceable research brief: the supplied facts, plausible candidates, governing text retrieved, unresolved questions, and the next human review step.
Use each public tool for what it actually does
| Tool | Use it to | Do not claim |
|---|---|---|
| search_hts_codes | Find current rows by keywords or prefix. | That the first result is the classification. |
| lookup_hts_code | Confirm a proposed code resolves and inspect its profile. | That existence proves product fit. |
| tandom_hts_hierarchy | See the parent and child context around a code. | That hierarchy alone resolves exclusions. |
| tandom_hts_notes | Retrieve applicable legal notes for review. | That the model interpreted every note correctly. |
| tandom_duty_calculate | Model duty and fee layers for supplied entry facts. | That a duty result validates the code. |
| tandom_adcvd_check | Screen code, product, origin, and parties for exposure. | That an HTS match determines legal scope. |
Before relying on a time-sensitive result, the model can also call tandom_regulatory_status to report the freshness signals returned by the public data service.
Prompt the model to produce a reviewable brief
Use this with an MCP-compatible client that has access to Tandom's public tool catalog. Replace every bracketed field. If a fact is unknown, write "unknown" rather than guessing.
You are helping a US importer research possible HTS classifications. You are not making the final classification. PRODUCT FACTS - Article: [plain-English name] - Function and end use: [facts] - Materials and composition: [facts or unknown] - Form, dimensions, capacity, or power source: [facts] - Imported as: [complete article / part / set / kit / unfinished] - Country of origin: [ISO-2] - Entry date: [YYYY-MM-DD] - Other relevant evidence: [specification links or facts] RESEARCH STEPS 1. List every missing fact that could change the heading. Do not infer it. 2. Use search_hts_codes to find plausible current rows. Search more than one factual phrase when necessary. 3. For each serious candidate, use lookup_hts_code and tandom_hts_hierarchy. A successful lookup proves existence only. 4. Use tandom_hts_notes for the relevant code or chapter. Identify the notes a human reviewer must interpret; do not claim they are resolved when the supplied facts are incomplete. 5. Do not invent CBP rulings, note text, legal citations, or 10-digit codes. 6. Do not label any candidate 'final,' 'approved,' or 'broker-grade.' OUTPUT - Supplied facts - Missing facts - Candidate codes with exact tool-returned descriptions - Relevant hierarchy and notes - Reasons each candidate may fit or fail - Conflicts and questions for the reviewer - Recommended next human review step Only after a reviewer chooses a candidate, offer to run duty and AD/CVD screening as separate analyses using the supplied origin and entry facts.
Review the brief before anyone relies on it
Confirm every cited code and text
Open the code in the Tandom catalog and the official USITC HTSUS. Confirm that quotations and descriptions came from tool output or the official source, not the model's memory.
Resolve missing facts with the supplier
If two candidates turn on composition, use, dimensions, or how the goods are imported, stop. Ask the supplier or engineering team for evidence. Do not ask the model to choose the "most likely" fact.
Search rulings outside the public Tandom catalog
The public MCP catalog does not include CBP CROSS search. A reviewer should search rulings.cbp.gov for materially similar products and check whether a ruling was modified or revoked.
Keep consequence analysis separate
Duty, PGA, Chapter 99, and AD/CVD results help the team understand exposure and information requirements. They should never be used as circular proof that the proposed code is correct.
Common pitfalls
- Calling tool access a guarantee. Current data does not prevent the model from misreading the facts or legal sequence.
- Treating search as classification. Search results are candidates, not decisions.
- Inventing confidence scores. The public reference tools do not return classification confidence.
- Assuming CROSS is built in. The public Tandom MCP catalog does not currently expose ruling search.
- Pasting secrets into chat. Configure the API key in the client's secure connection settings, never in the prompt.
- Sending confidential documents casually. Follow your organization's approved AI, privacy, and retention policies; redact unnecessary commercial details.