Menu
Skills Pricing What's MCP? Docs Methodology Support
Sign in

Connector Documentation

Proofwork provides certified AI skills as Model Context Protocol (MCP) servers. Connect once in your AI assistant of choice — Claude, ChatGPT, or any MCP client — and use the tools in natural conversation.

Overview

Every Proofwork skill is a narrow, benchmarked tool designed for high-stakes work where being approximately right isn't good enough. Each skill documents its data sources, methodology and known limitations, and runs on a frozen reference dataset — so you can audit its behavior instead of trusting a black box.

Available today

HTS tariff classification — tools for structured heading → subheading → leaf-code workflows, backed by CBP ruling precedents.

97.2% accuracy See how we measured this

Also on this connector, in beta

Restricted-party screening, post-classification trade flags (Section 301/232, agency requirements, AD/CVD exposure), and FDA medical-device classification. One connector, not one per skill — ask the assistant to list the skills.

Beta means no end-to-end run stands behind the skill yet. Two of the three do publish measured numbers — for how reliably the right record is surfaced, not for whether the assistant then picks it. Each skill's page states its sources, their as-of dates, and what the skill will not decide.

How to connect

Proofwork's MCP server lives at one public URL. Every supported client points at the same endpoint:

https://mcp.proofworklabs.com/mcp

Claude (web, desktop, or mobile)

  1. Open Customize → Connectors and select Add custom connector.
  2. Enter a name (for example Proofwork Labs) and paste the URL above. Leave the optional OAuth client fields empty.
  3. Select Connect and sign in on the proofworklabs.com page that opens.
  4. In a chat, enable the connector from the tools menu and ask your question.

On Team and Enterprise plans an organization owner adds the connector once for everyone under organization settings; each member then connects and signs in individually.

Claude Code

claude mcp add --transport http proofwork https://mcp.proofworklabs.com/mcp

Then run /mcp inside Claude Code, select the server, and follow the prompt to sign in from your browser.

ChatGPT

  1. Open Settings → Security and login and turn on Developer mode (availability depends on your plan and workspace policy).
  2. Go to chatgpt.com/plugins and select the plus button.
  3. Enter a name and description, paste the URL above as the MCP server URL, and create the connection.
  4. Sign in on the proofworklabs.com page that opens, then review the tools ChatGPT discovered.
  5. In a new chat, add the connection from the tools menu and ask your question.

Other MCP clients

Any client that supports remote MCP servers with OAuth or bearer-token auth will work. For programmatic integrations (Python MCP client library, curl, the Anthropic SDK with tool_use), see the Developer tab on the HTS get-started page.

Tools

Every tool is a read-only lookup against Proofwork's bundled reference data. Tools return reference text, precedents and a short selection guide; your assistant does the reasoning. Proofwork never calls an AI model, changes no data, and does not act on your behalf anywhere else.

Connecting adds every tool below. Two of them just describe the rest: list_skills names what is available and get_skill_guide returns one skill's workflow — so you can ask your assistant "what can Proofwork do?" rather than reading this page.

FDA Device Classification

Beta

Beta: no end-to-end run stands behind it yet. Its page states where the data comes from, its as-of dates, and what it will not decide. FDA Device Classification details.

  • find_fda_device_codes_payload — Find candidate FDA product codes for a device description: class, submission type, exemption and implant flags, FDA's definition, examples
  • get_fda_product_code_payload — Get the full record for one product code: 21 CFR identification text, premarket pathway, sibling codes, and what the record does not determine
  • find_fda_predicate_candidates_payload — List cleared 510(k)s under a product code as predicate candidates, with K-numbers and FDA database links — a research list, not an SE opinion

Trade Risk

Beta

Beta: no end-to-end run stands behind it yet. Its page states where the data comes from, its as-of dates, and what it will not decide. Trade Risk details.

  • screen_party_payload — Screen one party: candidate list entries with match tier, score breakdown, full evidence and triage instructions
  • get_party_record_payload — Get the complete list entry for a record id
  • screen_parties_batch_payload — Screen up to 50 counterparties in one call, compact results

Trade Core

Certified

Measured end to end against a published benchmark — how we measured it. Trade Core details.

  • identify_heading_payload — Find the 4-digit HTS heading from candidate evidence cards
  • select_subheading_payload — Pick the 6-digit subheading from group evidence
  • select_leaf_payload — Pick the exact 10-digit code from leaf codes + precedents
  • verify_code_payload — Compare selected code against sibling codes
  • extract_tariff_attributes — Extract tariff-critical attributes from a product description
  • get_suffix_tree — Browse the labeled code tree for any heading
  • explore_hts_heading — Inspect available codes under any heading

Post-Classification Trade Flags

Beta

Beta: no end-to-end run stands behind it yet. Its page states where the data comes from, its as-of dates, and what it will not decide. Post-Classification Trade Flags details.

  • check_trade_remedies_payload — Chapter 99 provisions whose U.S. note lists this code, with the note citation, the country named and the duty text
  • check_agency_requirements_payload — Partner government agencies whose published import requirements reach this code's chapter or heading
  • check_addcvd_exposure_payload — Antidumping and countervailing duty order notices whose published scope cites this code

Authentication & free tier

Proofwork uses standard OAuth 2.0 (Authorization Code + PKCE) for interactive clients and API keys for programmatic access. When you first connect, your AI assistant will open a sign-in page hosted at proofworklabs.com where you can sign in with email and password or with Google, and accept the Terms of Service. Proofwork asks for a single permission, mcp:tools (call the tools listed above). You are not prompted again on subsequent tool calls; when the authorization expires your assistant asks you to sign in again.

Free tier: 500 calls

Every organization gets 500 MCP tool calls for free, pooled across all Proofwork skills and all members. No credit card required. Plans are at proofworklabs.com/pricing.

If your company email domain matches an organization already on Proofwork, you'll auto-join it and share the company's subscription. No invite flow needed for corporate accounts. See our security page for full details on data handling, infrastructure, and compliance.

What data Proofwork receives

  • Received: only the arguments of each tool call — a product description and optional materials, intended use, country of origin, and HTS headings or codes.
  • Never received: your conversation, chat history, files, or your assistant's memory.
  • Stored: per call, your user ID, the tool name, a random request ID, success or failure, latency and time — for free-tier limits, billing and reliability. Tool inputs and results are not stored.
  • No AI model is called by Proofwork, and tool inputs are not used to train models.
  • Coverage measurement: off unless your organization switches it on. While on, each classification lookup also records whether a matching CBP ruling was found, how close the match was, and that ruling's public number — never the description, the code or who asked. Deleted after 90 days.

Full details, retention periods and service providers are in the privacy policy. To delete your account and data see support.

Worked examples

Three prompts for the tariff-classification skill, with the tool calls they produce. In practice your assistant orchestrates these itself.

The other three skills work the same way — ask in your own words and the assistant picks the tool. Screening a party: "Is Shenzhen Hua Ying Technology on any US restricted-party list?" · A settled code's downstream checks: "I classified this as 8413.70.2004, made in China — what extra duties or agency requirements apply?" · An FDA device: "What product code covers a single-use silicone laryngeal mask airway?"

1

Consumer product made of two materials

A precedent that ships in the reference data

Prompt:

"Classify a glass food storage container with a wooden lid and a wire bail-and-trigger closure."

Expected tool sequence:

  1. identify_heading_payload → picks heading 7013 (glassware of a kind used for table, kitchen, toilet, office or indoor decoration)
  2. select_subheading_payload → picks 7013.49 (table or kitchen glassware other than lead crystal or low-expansion glass)
  3. select_leaf_payload → picks 7013.49.2090 (valued not over $3 each, other)

Result: 7013.49.2090 — the glass body gives the article its essential character, so the wooden lid and the steel closure do not move it out of chapter 70. The tools return CBP ruling N327815 (2022-12-05) as the precedent, with the facts to confirm against your goods.

2

An incomplete article, with a country of origin

A precedent that ships in the reference data

Prompt:

"What's the HTS code for a blender base assembly from China, imported with or without the blade assembly and jar?"

Expected tool sequence:

  1. identify_heading_payload → heading 8509 (electromechanical domestic appliances with a self-contained electric motor)
  2. select_subheading_payload → 8509.40.00 (food grinders, processors and mixers; fruit or vegetable juice extractors)
  3. select_leaf_payload → picks 8509.40.0015 (blenders)

Result: 8509.40.0015 — the base has the essential character of the finished appliance, so it classifies as a blender rather than as a part. The country of origin drives duty treatment, not the code. Precedent returned: CBP ruling N343115 (2024-11-05).

3

A set of several articles sold together

A precedent that ships in the reference data

Prompt:

"Classify a yard maintenance set containing a battery lawn mower, a string trimmer and a blower."

Expected tool sequence:

  1. identify_heading_payload → returns candidates including 8433 (mowers for lawns, parks or sports grounds) and 8467 (tools for working in the hand, with self-contained motor)
  2. The set has no single heading of its own, so the classification turns on which article gives it its essential character — the mower
  3. select_subheading_payload → 8433.11.00 (powered, with the cutting device rotating in a horizontal plane)
  4. select_leaf_payload → picks 8433.11.0010 (electric mowers, including battery operated)
  5. verify_code_payload → compares the sibling leaves that split on power rating, to confirm the electric line is the right one

Result: 8433.11.0010. This is the case where the tools are most worth reading rather than trusting: a set can be classified as one article or as several, and the precedent returned — CBP ruling N330768 (2023-02-13) — is flagged with the facts your goods have to share for it to hold.

Want to know how it works? Every example above is a real CBP ruling that ships in the reference data, and none of them is a benchmark case — we do not publish the products we grade ourselves against. See the HTS skill deep-dive for how the skill works, where it is strong, where to verify, and our benchmark status.

Limits & availability

Tool response size

AI assistants cap the size of a tool result (Claude Code at 25,000 tokens, for example). Classification payloads are a few thousand tokens. Very large code trees (Chapter 99 headings such as 9903) are returned in parts: the tool lists the subheading prefixes you can request.

Tool timeouts

MCP clients enforce a timeout of a few minutes per tool call. Proofwork tools typically respond in under 2 seconds.

Hosting

Proofwork's MCP server runs on Google Cloud Run in us-central1 with managed TLS. Tool inputs and results are not stored — only the usage record described above, plus the coverage-measurement record if your organization switched that on.

Updates

Reference datasets (CBP rulings, HTSUS schedule) are refreshed on a published cadence. Version numbers and refresh dates are exposed in each skill's deep-dive page.

Troubleshooting

  • Cannot connect or sign-in fails: check the URL includes /mcp, leave optional OAuth client fields empty, allow pop-ups for proofworklabs.com, then remove and re-add the connector.
  • Asked to sign in again: the authorization expired; signing in again does not affect your account or usage.
  • Tools are not used: enable the connector for the conversation and name it in your request ("Use Proofwork to classify…").
  • "Free tier exhausted" error: the account has used its free calls; plan details are on the pricing page.
  • Invalid heading or code errors: headings are 4 digits (7013), subheadings 6 (7013.49), full codes 10 (7013.49.2090), and must exist in the HTS revision the tool reports.

More help, including how to disconnect, is on the support page.

Support

Questions, bug reports, or integration help:

Ready to connect?

Add Proofwork to your AI assistant in under a minute.

Get started with HTS