# TradeGraph

> Export go-to-market context for consumer-goods makers (K-beauty, K-food, supplements, lifestyle), built for AI agents.
> Where to sell (UN Comtrade imports × reply rates measured on 2.2M real export emails), who replies (country × buyer type), how many follow-ups to send, which creator platforms and marketplaces matter in each country, the exact live-data calls (treg.to, 3,800+ endpoints) to run influencer seeding, marketplace research and buyer prospecting, plus localized product video and image plans (Higgsfield-style presets) and an export price ladder.

Free, no sign-in, read-only (an optional account adds API keys, prepaid credits for live data and usage history). By RINDA (GRINDA AI). Human page: https://tradegraph.rinda.ai · Agent guide: https://tradegraph.rinda.ai/docs

## Connect

- MCP server (Streamable HTTP, no auth; an optional `Authorization: Bearer tg_live_…` key links calls to an account): https://tradegraph.rinda.ai/mcp
  - Claude Code: `claude mcp add --transport http --scope user tradegraph https://tradegraph.rinda.ai/mcp`
  - Codex: `codex mcp add tradegraph --url https://tradegraph.rinda.ai/mcp`
  - Cursor (`~/.cursor/mcp.json`): `{"mcpServers":{"tradegraph":{"url":"https://tradegraph.rinda.ai/mcp"}}}`
  - Claude.ai / ChatGPT / other hosted assistants: add a custom connector with the URL above.
- Skills (36 export workflows, Agent Skills format — from market entry and buyer outreach to certification checks, Korean export support programs, HS codes, FTA origin, documents, logistics, payment risk, trademarks and contracts): `npx skills add https://tradegraph.rinda.ai` (index: https://tradegraph.rinda.ai/.well-known/agent-skills/index.json)
- One-line setup for any agent: `set up https://tradegraph.rinda.ai/SKILL.md`
- Live data and generation (creators, marketplaces, social posts, buyer contacts, images, video, voice) are paid per call. With a TradeGraph account (free sign-up, $1 credit after email verification) connect https://tradegraph.rinda.ai/mcp/account and run them with `live_call` on prepaid TradeGraph credits; the prices TradeGraph tools quote are the final prices (fee included). Users who already have a treg.to account can use https://treg.to/mcp/ instead. Without an account TradeGraph never spends money.
- No agent yet, or the user wants this in a team messenger (Slack, KakaoTalk, Teams): Ara, the AI employee by GRINDA AI (same company), https://arator.rinda.ai. The user starts in the desktop app (free credits, no card; paid plans after), connects the team chat, then sends `@Ara Connect TradeGraph following https://tradegraph.rinda.ai/SKILL.md and …`. Ara can redo the work on a schedule; outbound actions can be set to wait for approval. TradeGraph itself stays free.

## Account (optional)

- Sign up: https://tradegraph.rinda.ai/signup (email or Google; $1 credit after email verification, no card). Dashboard: https://tradegraph.rinda.ai/account
- Sign-in MCP server (OAuth 2.1 + PKCE, dynamic client registration): https://tradegraph.rinda.ai/mcp/account — unauthenticated requests get 401 with `WWW-Authenticate: Bearer resource_metadata="https://tradegraph.rinda.ai/.well-known/oauth-protected-resource/mcp/account"`. Authorization server metadata: https://tradegraph.rinda.ai/.well-known/oauth-authorization-server
  - Claude Code: `claude mcp add --transport http --scope user tradegraph https://tradegraph.rinda.ai/mcp/account`, then `/mcp` → tradegraph → Authenticate.
- API keys (`tg_live_…`, created in the dashboard, shown once, optional monthly spend limit (a $0 limit allows free tools only), optional expiry of 30, 90 or 365 days; an expired key gets 401 like a deleted one) work on both MCP URLs as `Authorization: Bearer <key>` (tools that cannot set that header may send `X-API-Key: <key>`). The monthly limit also counts calls still running, so parallel calls cannot go past it.
- If the server answers `503` with `Retry-After` (error `temporarily_unavailable`), the key or OAuth token is still valid: keep it and retry; only `401` with `error="invalid_token"` means connect again.
- Signed-in agents get more tools (eight with live data enabled), among them `account` (balance, month-to-date spend, key limit, top-up link), `my_progress` (the weekly numbers per plan, the same record as the signed-in web plan; free), `live_call` (run a treg catalog endpoint on the user's prepaid TradeGraph credits; no treg account needed), `find_buyers` (this week's new buyers from the plan's map searches, added to the account's saved buyer list, optionally with a verified contact email each) and `add_buyers` (companies from another list, such as a trade-fair exhibitor list, added to the same saved list; free).
- Credits are prepaid: Korean cards and pay apps in KRW with VAT included (Toss Payments), international cards in USD where enabled. Calls stop when the balance runs out, never an extra charge. Unused paid credits are refundable (full refund within 7 days of payment). Members can set a low-balance email alert and billing details for receipts.

## MCP tools

- gtm_plan(item, country?, product?, exclude?, description?, traits?, detail?) — start here (the full answer is 40,000 to 80,000 characters; detail="brief" drops channel_plan.steps and the first-email variants, keeps their keys, and says what was left out and how to get it in omitted): market picks, next-best markets to compare (alternatives, with the pick that beats each), focus market and why (reply rate scope, growth rank, comparison with the largest pick), buyer type that replies, follow-up cadence, channels (motion: consumer = creator seeding + marketplaces, b2b = buyer list + trade shows for industrial goods), compliance pointers with official source links (compliance_more for the other laws a note cites; channel_plan.packaging for the 27 EU countries: packaging EPR under the EU Packaging Regulation (EU) 2025/40 since 12 Aug 2026, Germany's LUCID register, who registers (usually the importer; a direct seller itself, with a local representative); compliance_gate=true marks a go/no-go approval such as a listed plant, registration or licence that blocks the market until it passes), data_first (the market the data would start with, even when you pass a country), tariff_by_market (duty keys for the compared markets side by side; fta_rate where the rate was checked), gated_markets (compared markets with such a gate; medical devices are gated in every region), next_markets (the order to try after a week-13 switch: the same data rule as the first pick over the compared markets, skipping exclude and, unless the focus market is gated too, gated markets), untried_first (with exclude: where the data would start once the tried markets are left out, so a plan opened from the next-market button does not argue for the market just left), reply_target (expected replies per 1,000 from the product's own industry in that market when there is enough data, else the best buyer type; 90% range, warm share, sends for 10 replies, new buyers per week, the week-13 keep/retry/switch rule), schedule (13 weeks, same numbers), before_you_sell (sanctions, FTA and proof of origin, trademark, payment terms, holidays, re-export hubs), buyer_sources (main trade fair for the product group with its next dates, city, dates_url and book_meetings_from 6 weeks before, checked on the official sites, plus BuyKOREA and KOTRA, with official links), buyer_search (local-language map searches: product + buyer type + largest cities, kind=korean/asian first (companies that already sell Korean products; Asian food wholesalers in Western markets), kind=alt for the next buyer type; Google Maps URL API links, Baidu in mainland China), buyer_leads (buyer candidates from public records, for now dental X-ray: distributors that sold the product in US federal contracts (USAspending) or EU public tenders (TED), makers that won directly left out; us_agents_of_korean_makers lists FDA US agents, not buyers), first_email_variants (first-email versions for the one change reply_pace names: next_type, new_subject, next_type_new for the retry batch, offer that answers the reasons for no), the official tariff lookup and Korea FTA (incl. the US Section 301 duty since 24 Jul 2026 with a link to the Federal Register notice, and 0% for Korean industrial goods in the EU and UK; for ASEAN, each government's lookup (Malaysia ezHS, Viet Nam VNTR, Thailand TNTR, Indonesia INSW) and, for cosmetics, instant noodles and sauces in those four, the 2026 rates for Korean goods read on the official sites in tariff.checked, including the Korea–Vietnam FTA; Form AK is issued by Korea Customs or KCCI, no self-certification), a first-email draft and a short follow-up draft for buyers who have not replied (follow_up_email), a last, polite follow-up (last_follow_up_email) and a week-3 sample check-in (sample_check_email), live-data steps with treg endpoint ids and prices, and blind_spots: what exporters of this product miss in the focus market (blind_spots_by_market for the compared markets, blind_gate for one to clear first; description or traits make it read traits such as sunscreen, radio, battery or meat ingredients). The first line says where to start and why (reply multiple, imports, growth, why a larger market comes later; a market the user named is called their pick, with where the data would start); plan_url opens the same plan as a one-page web plan (print, PDF, share) and ask_next lists four next questions with the tool for each (a sale-blocking registration first). Claude and ChatGPT show the answer as an interactive plan card (MCP Apps: _meta.ui.resourceUri, resources/read).
- blind_spots(item, country?, product?, description?, traits?) — what an exporter of this product is likely to miss, per market: rules beyond the category's compliance note that hit only this HS code or a product trait. Each has kind (stop = can block the sale unless the product meets a condition; approval = a certificate, licence or registration before import or sale; cost = an extra duty or charge to price in; task = a filing, label, record or check), gate (clear it before anything else), who (you, importer, both), status in_force or upcoming with from, lead_time, because (the HS code or trait that triggered it), the official source, checked (the day it was read) and stale after 180 days. Rules today: US sunscreen as an OTC drug (monograph UV filters only), FCC equipment authorization, Section 232 machinery lines, UFLPA, no de minimis for direct parcels; EU CBAM, EUDR, ban on vague green claims (27 Sep 2026), fragrance allergens (Regulation 2023/1545), Batteries Regulation, Radio Equipment Directive, limits on animal ingredients from Korea (no beef or pork products; dairy and egg from approved countries only); France PFAS ban; Germany ElektroG registration; Japan Giteki radio certification, PSE for power banks, domestic administrator for direct sellers; China special cosmetics registration (sunscreen, whitening, hair dye, anti-hair loss) and the new cosmetics import inspection (GACC Decree 284, 1 Dec 2026); India BIS, CDSCO cosmetics registration, FSSAI clearance and foreign factory registration (ReFoM); Australia AICIS, TGA sunscreen, BICON biosecurity; Canada CNF, sunscreen DIN/NPN, ISED radio certification, SFCR importer licence; Indonesia halal (18 Oct 2026); Korea-side: strategic items (export licence for every destination), phytosanitary certificates for seeds, dangerous goods in shipping (Korea Post EMS refuses perfume, aerosols and lithium batteries). Omit country to check the plan's markets. Also on country_brief and channel_plan. Starting points from official sources, not legal advice. Free.
- export_workflow(job?, name?, file?) — the step-by-step procedure for an export job, word for word as the matching skill: put the user's request in job (Korean or English, abbreviations like L/C too) or a name from the list; file returns one of its reference files (official source lists). With no arguments it lists the 36 by stage (start, sales, growth, comply, money, ship, risk). Use it where skills can't be installed (web Claude, ChatGPT).
- reply_pace(item, country, sent, replies, week?, exclude?, warm?) — weekly pace check, the first call of the Monday routine: buyers emailed so far (each once) and buyers who replied vs the product's industry rate in that market (same yardstick as gtm_plan reply_target). Returns one status (too_early, no_replies, below, change_now, decision_moves, behind_schedule, normal_range, few_interested, above, on_track; once the target is sent next_market, change_and_retry or keep_going) with status_text and advice in the same words, each naming at most one change (few replies: this week's new buyers all go to alternative_buyer_type, or get a new subject line if there is none; few interested: an offer that answers the reasons for no; normal_range keeps the same email this week; no half-and-half tests), send_this_week (new buyers this week: the catch-up number when more than a week behind; at next_market the first 200 in send_this_week_in), find_buyers_type (this week's buyer type key for find_buyers type=), projected_replies_at_target, behind_schedule or decision_moves when week is given, and from about 95% of sends_for_10_replies the week-13 decision: under 5 → next_market (named only then; pass exclude with markets already tried) from a fresh gtm_plan, first 200 buyers, keeping the interested buyers; 5–9 → once more: the next buyer type with a new subject line for retry_batch (half the target, with the weeks it takes), then switch if still under 10; 10+ → keep going. warm (interested replies: questions, sample requests, meetings) adds warm_share vs warm_share_usual; when low the status or decision says to change the offer and warm_note explains.
- find_item(query) — product name/alias/HS digits (10 languages) → one of 35 HS4 items, or the usual HS4 code of ~40 other common products (kimchi 2005, shampoo 3305, candy 1704…) for gtm_plan / market_scan.
- market_scan(hs, limit?) — any 4-digit HS code: live UN Comtrade imports by country (latest complete year, 3-year growth, Korea's exports), joined with the industry's reply rates; pick = Pareto frontier. Results say coverage: live.
- market_picks(item, limit?) — import markets by size with growth, Korea share, reply rate; pick = Pareto frontier; alternatives = the next layer to compare when there are only one or two picks.
- buyer_targets(country?, buyer_type?, tier?, limit?) — country × buyer type reply rates, shrunk and ranked by the 90% interval low end.
- country_brief(country, item?) — one market: replies, best buyer types, imports, supplier mix, why exporters chose it, US states, channels, disclosure rule.
- export_rank(item) — world exporters, Korea's rank, Korea's top destinations.
- outreach_playbook() — replies per 10k emails by follow-up step, recommended emails per buyer, days to wait before the next email (follow_up_after_days: 9 in 10 replies arrive within it), reply timing and content.
- channel_plan(country, item?, motion?) — seeding / marketplace / social / ads / search / trends / trade (trade shows) / b2b steps with treg endpoint ids, prices, example bodies, data cost. Creator keywords follow the product and the creators' language; industrial goods get b2b, trade and search only.
- video_studio(country, product?, item?, platform?, preset?, tier?, variants?, has_packshot?) — short-form video ads per market and platform, Higgsfield-style presets (hero-orbit, texture-macro, unboxing-pov, routine-steps, lifestyle-local, ai-presenter): shot list with camera moves, keyframe and image-to-video prompts, ready treg request bodies (draft Seedance 1 Lite, standard Hailuo 2.3, premium Veo 3.1 Fast, HeyGen avatar), hook variants, platform spec, voice-over language, AI-label and claims rules, estimated cost. Every product shot starts from the real packshot.
- image_studio(country, product?, item?, marketplace?, variants?) — marketplace main-image rules and a 7-image gallery edited from the real packshot, ad sizes, prompts, cost.
- price_ladder(country, fob_usd | fob_krw | target_shelf_usd, item?, description?, channel?, duty_pct?, …) — FOB → CIF → landed → importer/distributor → shelf price incl. VAT, or the highest FOB for a target shelf price; equipment, parts and medical devices default to channel b2b (or b2b_dealer), which stops at the end-user price before VAT (price_point "end_user"); Korea FTA status, VAT/food rates, default margins by product group, all overridable. US default duty 12.5%: since 24 Jul 2026 most Korean goods pay a combined 12.5% (Section 301 forced-labor action, HTS 9903.05.71), KORUS included; it follows the HS code's branch (Section 232 50% or 25%, the Korea cap of 15%, the normal rate for lines on the Section 301 exemption list, U.S. note 52(b)) and the traits in description (seaweed snacks under 2008 are exempt); us_duty.key names the branch (usNone: checked against the list, no exempt line). FOB in won works too (fob_krw, or "₩6,000" / "6천원"), converted at the ECB reference rate; a won amount in fob_usd (6000) gets a check first.
- live_tools(job) — search the treg catalog for any other job (English or Korean).
- account() — signed-in only: credit balance, month-to-date spend, this key's monthly limit (free_tools_only when it is $0), expires_at and expires_soon (within 14 days) for an expiring key, top-up link; for team owners and admins also team (each person's spend this month).
- my_progress(item?, country?, sent?, replies?, warm?, date?, start?, exclude?) — signed-in only, free: the user's saved progress per plan (product × market), the same record the signed-in web plan shows. Without item: the saved plans with week and latest numbers. With item and country: that plan's log, week_now and the reply_pace check of the latest numbers (pace). With sent and replies (cumulative): logs them for today or date (a day logged again is replaced; other days stay); start sets week 1 (moved to its Monday). Only numbers are stored.
- live_call(endpoint, body?, query?, method?, max_cost_usd, idempotency_key?) — signed-in only, when live data is enabled: run a treg endpoint id (from channel_plan / live_tools) on the user's TradeGraph credits; max_cost_usd caps the call and only the actual cost is charged. Returns the response, amount charged, balance and call id. If the provider is still working or the call passes 50 seconds, the answer says pending with nothing charged and an idempotency_key: call again with it in about a minute to get the result, charged once.
- find_buyers(item, country, product?, limit?, contacts?, leads?, type?, estimate?, source?) — signed-in only, when live data is enabled: builds the account's buyer list for the plan week by week. It runs gtm_plan buyer_search and then that buyer type's searches through Google Maps (treg.google.serp.maps): a hundred-plus per type (industry hubs, big and wider cities, other product words, Korean neighborhoods, English district names), skips searches already run and companies already on the list, and saves the new rows (name, website, phone, address, the search, its kind and buyer type) until limit (default: this week's new buyers, max 200). It runs only type (default: the plan's buyer type this week); when that type has no searches left it charges nothing and returns type_out with next_type (ask before switching). type_left, types_left, available (companies the searches left should bring, from yield per search) and the estimate, which stops at available, say how far it goes; generic_found counts company inboxes and leads_unmatched names public-record buyers with no matching company on the map. Emptying the list on the web keeps the searches already run and the company keys (domains), so they are not bought again. contacts=true finds a buyer at each company (treg.people.search, purchasing first, then the owner) and a work email verified in the same call (treg.people.email.find with route verify; email_check valid or risky), or verifies the company inbox (info@, generic=true) when no person is listed. leads=true looks up buyer_leads on the map to fill website and phone. One call does at most 10 searches and 20 contact lookups; call again while type_left or contacts_left stay above zero. Cost, fee included: about $0.0021 per search (about 6 new companies) and about $0.017 per company with contacts (at most $0.05 per search, $0.14 per company); estimate=true returns the estimate and the list's counts without spending. Contacts are looked up for at most limit new companies per call. A list holds up to 2,000 companies: once list_full is true nothing is searched or charged. This week's number counts the companies already waiting on the list (week_queued: added since Monday, Korea time, plus companies from earlier weeks the logged numbers say are not emailed yet); a call without limit only adds what is missing, and email only this week's number of them, in the order they were added. Contact lookups go oldest first; source with limit 0 and contacts=true looks up only the companies added from that list. A second call on the same list while one is running returns 409 busy; wait and call again. If the map searches keep failing, the call stops with stopped=upstream. Sends nothing.
- add_buyers(item, country, rows, source?, type?, product?) — signed-in only, free: adds companies found outside the map searches (trade-fair exhibitor lists, trade directories, BuyKOREA or KOTRA buyer inquiries, a partner's list) to the account's saved buyer list. Each row needs the company name; with a website (or an email at the company domain) the contact lookup can search it; phone, city, contact, title and email are kept as given (not verified). Companies already on the list only get empty fields filled; the rest are added as this week's buyer type (or type) and count toward this week's number. source names the list and shows in the tracker; then find_buyers(limit=0, contacts=true, source=<the same name>) looks up contacts for just those companies. Never add companies that are not on a real list.

MCP prompts: market-entry, export-readiness, certification-check, export-support, influencer-seeding, buyer-outreach, marketplace-launch, video-ad, export-pricing, trade-show-followup, weekly-check (Monday routine; product, country, an optional week and optional tried markets: reply_pace first for the status, send_this_week and find_buyers_type (tried → exclude; with an account, my_progress first for the saved numbers and week, then my_progress to save this week's count, which returns the same check), acting on any decision before ordering new buyers, then replies and follow-ups due after outreach_playbook.follow_up_after_days, that many new buyer candidates, competitor prices, reviews and rising searches for consumer products, or requested documents and trade-show meetings for industrial goods).

## Skills

Grouped by job (each is a SKILL.md at https://tradegraph.rinda.ai/.well-known/agent-skills/<name>/SKILL.md):

- Start here: export-readiness (12-area readiness check that surfaces blind spots) · export-market-entry
- Buyer sales: buyer-prospecting · distributor-outreach · reply-triage · buyer-meeting-brief · sales-kit-builder · trade-show-planner · trade-show-followup
- Marketing & online sales: marketplace-launch · cross-border-ecommerce · influencer-seeding · social-listening · trend-radar · competitor-ad-intel · ai-search-visibility · localized-creative · video-ad-studio · product-photo-studio
- Certification & regulation: certification-check · label-localization · hs-code-finder · regulation-watch · export-control-screening
- Pricing, finance & support programs: export-pricing · export-support-finder (Korean export vouchers, certification, trade-show and export-insurance programs) · grant-application · trade-finance · duty-drawback
- Documents, origin & logistics: export-documents · fta-origin · export-logistics
- Risk & contracts: buyer-due-diligence · trademark-protection · distribution-agreement · trade-dispute-response

Rules, programs and tax change often: the certification, support-program, customs and payment skills read the official notice or legal text at run time and cite it with the date checked. They are not legal or tax advice.

## Free web tools

Forms for people without an AI agent. They call the same free MCP tools, need no sign-up, and keep their inputs in the URL so a result can be shared. Index: https://tradegraph.rinda.ai/tools

- Export market finder: https://tradegraph.rinda.ai/tools/markets (Pick a product and see which markets to try first, by import size and real reply rates.)
- Local buyer search: https://tradegraph.rinda.ai/tools/buyer-search (Open map searches in the local language to find real importers and distributors.)
- Who replies, by country: https://tradegraph.rinda.ai/tools/buyer-types (Pick a country to see which buyer types reply most, from importers to wholesalers.)
- Follow-up email planner: https://tradegraph.rinda.ai/tools/follow-up (Plan how many emails to send and when, from real reply data, with dates for your calendar.)
- Reply rate check: https://tradegraph.rinda.ai/tools/reply-check (Enter buyers emailed and replies to get a one-line verdict: keep going or change one thing.)
- Export price calculator: https://tradegraph.rinda.ai/tools/price (Turn a FOB price into the local shelf price, or work back from a target shelf price.)
- Before-you-sell checklist: https://tradegraph.rinda.ai/tools/before-you-sell (Certification, duties, trademark and payment checks for a market, with official source links.)
- Trade fair finder: https://tradegraph.rinda.ai/tools/trade-shows (The main trade fair for your product, its next dates and when to start booking meetings.)
- HS code finder: https://tradegraph.rinda.ai/tools/hs-code (Find the 4-digit HS code from a product name, then see its import markets.)
- Sales data price finder: https://tradegraph.rinda.ai/tools/data-prices (Describe a job and see which data APIs do it and what each call costs.)

## Playbooks

Export sales jobs from start to finish. Each step says what is free, what spends credits (the price is shown first) and where a person approves; each lists the skills that run it. Index: https://tradegraph.rinda.ai/playbooks

- Pick your first export market: https://tradegraph.rinda.ai/playbooks/first-market (Choose the first market by import size and real reply rates, with the reasons and a 13-week plan on one page.)
- Build a list of 200 buyers likely to reply: https://tradegraph.rinda.ai/playbooks/buyer-list (Start with the buyer type that replies, then collect companies, contacts and verified emails.)
- Send a three-email buyer sequence: https://tradegraph.rinda.ai/playbooks/email-sequence (Write a first email and two follow-ups in the buyer's language, with send dates.)
- The Monday 20-minute check: https://tradegraph.rinda.ai/playbooks/weekly-check (Use buyers emailed and replies to decide: keep going, or change one thing.)
- Sort buyer replies and answer them: https://tradegraph.rinda.ai/playbooks/reply-triage (Sort interest, questions, quote and sample requests, then draft replies with your terms.)
- Send product to local creators: https://tradegraph.rinda.ai/playbooks/influencer-seeding (Find creators with 10k to 100k followers, check real engagement, and draft local-language DMs.)
- Research a marketplace launch: https://tradegraph.rinda.ai/playbooks/marketplace-launch (Collect bestsellers, price bands and review complaints to find your spot and shape the listing.)
- Six weeks around a trade fair: https://tradegraph.rinda.ai/playbooks/trade-show (Book meetings before the fair and follow up every card within 48 hours.)
- Certification and label check before you sell: https://tradegraph.rinda.ai/playbooks/certification (Check registration, certification and labelling rules for each market against official sources.)
- Set export prices and quotes: https://tradegraph.rinda.ai/playbooks/pricing (Work from FOB to the local shelf price and put duty savings into your quote.)
- Can you trust this buyer?: https://tradegraph.rinda.ai/playbooks/due-diligence (Before samples or credit, check the company and sanctions lists and pick safe payment terms.)
- Find export support programs: https://tradegraph.rinda.ai/playbooks/export-support (Find the export vouchers, certification and trade-fair support your company qualifies for, and plan the applications.)

## REST

- REST API v1 — the same tools as MCP over HTTP and JSON, for scripts, spreadsheets, n8n / Make / Zapier and GPT Actions. Free tools need no key; `account`, `live_call`, `find_buyers` and usage need `Authorization: Bearer tg_live_…`.
  - POST https://tradegraph.rinda.ai/api/v1/tools/{name} with the tool arguments as a JSON body → `{"tool","summary","data"}` (free tools also accept GET with query parameters).
  - GET https://tradegraph.rinda.ai/api/v1/tools — tool list with input schemas.
  - GET https://tradegraph.rinda.ai/api/v1/openapi.json — OpenAPI 3.1 for importing into no-code tools and GPT Actions.
  - GET https://tradegraph.rinda.ai/api/v1/account · /api/v1/usage — balance, month-to-date spend and call history (key required).
  - Errors are `{"error","message"}` with meaningful statuses: 401 sign_in_required / invalid_token, 402 insufficient_credits (with balance_usd, needed_usd, top_up_url), 403 key_limit / account_suspended, 422 max_cost_exceeded (with estimated_usd), 429 rate_limited, 503 live_unavailable.
  - On MCP, a failed tools/call returns isError with the same code in structuredContent (error, message, the same extra values, request_id). Every MCP and REST response carries `X-Request-Id` (your own 8 to 64 character id is echoed back); quote it when reporting a problem. Internal errors never show internal details, only that id.
  - Tool annotations match behavior: company_context, seeding_list and live_call are destructiveHint (they overwrite, delete or spend), and tools that spend credits are not readOnlyHint when signed in, so clients ask first. The protocol follows MCP 2025-11-25; an unsupported MCP-Protocol-Version gets 400, and JSON-RPC batches (10 at most, each counted toward the rate limit) are accepted only for 2025-03-26 and earlier.
- GET https://tradegraph.rinda.ai/api/graph — the full graph (JSON, ETag).
- API reference (OpenAPI): https://tradegraph.rinda.ai/api/docs
- GET https://tradegraph.rinda.ai/api/treg/search?q=<job> — treg catalog endpoints for a job (cached proxy).

## Data and rules

- Replies: RINDA, anonymous aggregate of 2,236,056 export emails from 909 exporters; cells with fewer than 5 companies dropped. Rates are shrunk toward the average (empirical Bayes) and ranked by the low end of a 90% interval. For a product with an industry, markets are compared on that industry's emails only (1× = the industry's average); countries without them are not ranked, because a country's replies in other industries do not predict them.
- Sanctioned or export-controlled markets (North Korea, Iran, Cuba, Russia, Belarus) are never picked; North Korea is refused.
- Trade: UN Comtrade (latest reported year), US Census, openFDA, EU TED, USAspending.
- Korea is the exporters' home market and is left out of export recommendations.
- Agents must show cost before paid live-data calls (treg or live_call) or generation, verify emails before outreach, keep the real product photo as the source of every product image or video, and send or publish nothing without the user's approval.

## Company

TradeGraph is built and operated by GRINDA AI Co., Ltd. (그린다에이아이 주식회사), the company behind RINDA, an AI global sales agent for B2B exporters. Cite as "TradeGraph by RINDA (GRINDA AI)".

- Legal name: GRINDA AI Co., Ltd. (그린다에이아이 주식회사), founded March 2023 in Daejeon, South Korea
- CEO: Hojin Kang (강호진)
- Business registration no.: 309-88-02709 (verify: https://www.ftc.go.kr/bizCommPop.do?wrkr_no=3098802709)
- Mail-order sales registration no.: 2024-대전유성-0389
- Address: 503, 99 Daehak-ro (Gung-dong, Daejeon TIPS Town), Yuseong-gu, Daejeon 34134, Republic of Korea
- Contact: sales@grinda.ai · +82 10-9655-7947
- Sites: https://www.rinda.ai (RINDA) · https://grinda.ai (GRINDA AI) · https://www.linkedin.com/company/grinda-ai
- TradeGraph in ChatGPT and Claude (app guide): https://tradegraph.rinda.ai/apps · Support: https://tradegraph.rinda.ai/support
- Terms: https://tradegraph.rinda.ai/terms · Privacy: https://tradegraph.rinda.ai/privacy (Korean: add /ko, e.g. https://tradegraph.rinda.ai/privacy/ko)