TYDI — The Trade Me AI-First Operating System

0 The 60-second answer to your four questions

1. What does the Trade Me API look like?

A single official REST API (JSON/XML), OAuth 1.0a, capped at 1,000 authenticated calls per user per hour. It exposes the entire per-listing lifecycle: create, photo, edit, relist, price, answer buyer questions, mark shipped, leave feedback, and scan competitor pricing. Genuinely capable.

2. What are the automations?

Everything a seller does to a single listing is an API call. The whole sell→fulfil→feedback loop can run unattended. What the API won't do is bulk-feed thousands of SKUs in one shot, or auto-fill electronics specs from a model number — and both gaps are exactly where AI earns its keep.

3. How automated can we make it?

The store already lists ~16,600 items every 2 months (~277/day) — comfortably under the rate limit. You can drive 90%+ of the listing + service workload through software, with a human approving batches and handling edge cases. Refurb's "every unit is unique" nature makes this easier, not harder.

4. Can we run the Trade Me side through AI super-efficiently?

Yes — and the strategic kicker: no Trade Me seller has built this. eBay has a deep third-party SaaS ecosystem (Snap2List, 3Dsellers, Repricer, eDesk); Trade Me has almost none. So on eBay this is table stakes; on Trade Me it's a durable operational moat you build once and own.

The wedge: a custom photo/IMEI → live-listing pipeline + LLM customer-service layer, running direct on the Trade Me API on your existing Node/Fastify/Postgres/Railway stack. That's the whole game.

1 What the Trade Me API actually is

Dimension	Reality	Implication
Protocol	REST, JSON or XML, base `api.trademe.co.nz`; sandbox `api.tmsandbox.co.nz`	Standard build; sandbox is auto-approved instantly for dev
Auth	OAuth 1.0a (3-legged, HMAC-SHA1). Not OAuth 2.0.	Slightly fiddly signing, but solved. Tokens ~permanent (die after 6 months unused / on password change)
Scopes	`MyTradeMeRead`, `MyTradeMeWrite`, `BiddingAndBuying`. Selling needs MyTradeMeWrite	One write token runs the whole seller operation
Rate limit	1,000 authenticated calls / user / clock-hour. Catalogue GETs are exempt	Plenty for 277 listings/day. Higher limits "at Trade Me's discretion"
Bulk feed	Discontinued. Old "Feeds / My Products" path is dead → Trade Me redirects bulk sellers to Tradevine	No single-call batch upload. You loop one `POST` per listing — fine at our volume
Electronics catalogue prefill	Doesn't exist. Catalogue match by model/barcode only covers DVD/Blu-ray	The model→spec mapping is your AI step. A feature gap we turn into an edge

The endpoints that matter (the seller's toolkit)

Create / validate listing — POST v1/Selling (+ POST v1/Selling/Validate dry-run). Title ≤80, description ≤2048, up to 20 photos, SKU field, ProductSpecification (GTIN/Brand/ManufacturerCode), Attributes (Brand/Model/Condition).
Photos — POST v1/photos/Add (base64 → returns PhotoId), then attach IDs.
Edit / withdraw / relist — POST v1/Selling/Edit, /Withdraw, /Relist, /RelistWithEdits (relist and reprice in one call).
Drafts — POST v1/Selling/Drafts; closest thing to "sell similar / templates" — one draft per product line, cloned programmatically.
Category & attribute lookup (rate-limit-exempt) — GET v1/categories/{cat}/attributes tells you which structured fields a category needs; attributesuggestions resolves messy model fragments.
Sold / unsold / live items — GET v1/mytrademe/solditems, /unsolditems, /sellingitems. Sold items carry buyer, shipping, status, tracking.
Fulfilment (writable) — POST v1/MyTradeMe/Status/{purchaseId}/GoodsShipped, SendPaymentInstructions, CourierParcels, PayNowRefund.
Buyer questions — GET v1/listings/questions/unansweredquestions (account-wide) → answerquestion (≤500 chars). The auto-responder hook.
Feedback — read + POST v1/mytrademe/feedback, /Feedback/Reply. Auto-leave positive on SaleCompleted.
Fees (margin-aware) — POST v1/Selling/Fees, /SuccessFees, /RelistFees preview cost before you commit — so the bot never lists at a loss.
Search / pricing intel — Search/General?member_listing={competitorId} returns a competitor's live listings; Search/PriceHistogram gives price distribution. Internal use only (§9 T&Cs).

2 Two 2026 rule changes — deal-checklist items, not engineering details

1. From 10 April 2026, new Marketplace API registration is in-trade-sellers-only. TYDI qualifies as a registered in-trade business — but the new owner must register the production app under their own in-trade selling account.

2. API app registrations are NON-transferable between members. You cannot inherit the vendor's existing API credentials — post-settlement you re-register the app fresh under the new ownership entity.

Action: add to the DD request — confirm the in-trade selling account + 19-year / 99.16%-feedback reputation transfers with the business (it's one of the most valuable assets in the deal), and budget for a clean API re-registration on day one. Neither blocks the build; both bite if discovered late.

3 Automatable vs Manual — the clean split

✅ Automatable unattended (API)	❌ Still manual / web UI / human
Create listings (auction or Buy-Now), validated first	True bulk-feed sync at scale (→ Tradevine, or just loop)
Upload + attach up to 20 photos	Store/Storefront branding & layout setup
AI enrichment of Brand/Model/Condition/GTIN	Electronics catalogue prefill (doesn't exist → our AI fills it)
Edit, withdraw, relist, relist-with-reprice	Native scheduled auto-relist (orchestrate it yourself)
Detect unsold items → reprice → relist	Adding photos to a populated live listing without a full Edit
Mark shipped, send payment instructions, attach tracking	Free-form private buyer DMs (no inbox API — only Q&A + enquiry email)
Read + auto-answer buyer questions (≤500 chars)	Production app approval + in-trade verification (one-time)
Auto-place + auto-reply to feedback	Anything caching/redistributing Trade Me data (T&Cs)
Preview every fee before acting · scan competitor pricing

Read it this way: the per-listing lifecycle is first-class automatable. The gaps are (a) bulk plumbing and (b) storefront chrome — neither of which is where the money is for a unique-unit refurb business.

4 What the best eBay refurb sellers do — and what ports to Trade Me

The eBay world runs on a deep SaaS layer (3Dsellers, Linnworks, Repricer, eDesk, Snap2List, ChannelAdvisor). Almost none of it integrates with Trade Me. So we steal the plays, not the tools. The 80/20 of elite operators is three moves:

Industrialise photo → listing. Never type a title. Phone snap + model/IMEI → AI draft → bulk publish. ~30s/listing vs minutes. ~80% of the time saved.
Certify devices (IMEI/serial) and bake it into price. A test report auto-fills specs and justifies a 5–10% premium.
Obsess over service metrics. Auto tracking, on-time dispatch, fast CS, auto-feedback → higher rating → better placement → more sales at the same spend.

The point: the plays that don't port (comp-scraping ML repricers, eBay's Refurbished Program) were the wrong plays for unique-unit refurb anyway. The 80/20 ports almost perfectly — and because no TM competitor has built it, it's a moat, not parity.

Portability verdict

Play	Ports?	How
Photo/model → listing generation	✅ Fully — biggest edge	PhotoRoom API (~$0.02/img) + vision LLM → TM attributes → `POST v1/Selling`. No Snap2List for TM = moat
AI photo cleanup at scale	✅ Drop-in	PhotoRoom/Pixelcut APIs are marketplace-agnostic
IMEI/serial cert → spec + premium	✅ High value	Device-level, marketplace-agnostic. NZ trust + price lever
Unique-unit SKU + FIFO costing	✅	Build in Postgres, or Tradevine (native TM FIFO) as backbone
CS auto-responder / FAQ deflection	✅ Build it	LLM over `unansweredquestions`. The 6–8 refurb FAQs are identical in NZ
Order/dispatch/tracking	✅ Build, medium	API + NZ Post/Aramex. Logic is simple
Dynamic repricing	⚠️ Lightweight only	No sold-comp feed. Build age-based markdown — the right strategy for unique stock anyway
Comp/sold-price intelligence	⚠️ Thin	Monitor competitor asks via Search; sold-price depth isn't there
Native auto-positive-feedback	❌ (no TM equivalent)	Build auto-request + negative-alert version
The entire eBay SaaS ecosystem	❌	Doesn't integrate. Build, don't buy (Tradevine the one exception)

5 The AI-first architecture (on your stack)

Build direct on the Trade Me API, on Node/Fastify/Postgres/Prisma/Railway.

Build direct, or use Tradevine? The "no bulk API" gap barely applies to refurb — bulk feeds matter when you sync one SKU across many units; refurb is qty-1 unique units, so the per-listing loop is the natural model. Sanity check: 277 listings/day × ~6 calls ≈ ~70/hour vs the 1,000/hour ceiling. Comfortable. Recommendation: build direct — keep Tradevine as a known fallback only if stock accounting gets hairy.

Intake

Bench operator photographs unit + scans IMEI/serial

Uploads to an intake queue (simple web form, phone camera)

▼

Enrich — the crown jewel (the AI step Trade Me doesn't do)

Photos + specs → a complete draft listing

PhotoRoom API → clean white-bg images (~$0.02 each)
IMEI/serial lookup → device identity + base specs
Vision LLM → Brand→Item→Model title (≤80, leads "Refurbished"), structured attributes, conversion subtitle, clean description w/ condition + 12mo warranty, suggested grade (A/B/C)
Price engine: cost-basis + margin floor, age-markdown rule, competitor-ask sanity check

▼

Approve — human-in-the-loop, batch not per-item

Reviewer sees a grid of drafts, edits inline, ticks approve

200 listings reviewed in minutes; Selling/Validate runs on every row first

▼

Publish

POST v1/Selling (+ photo upload + fee preview)

SKU written back to the Postgres ledger (grade + serial + intake date)

▼

Operate — always-on workers on Railway

The loop runs itself

Q&A bot: poll unansweredquestions → LLM draft → auto-send simple FAQs, queue judgment calls
Fulfilment: poll solditems → payment instr → tracking → mark GoodsShipped → notify
Relist/reprice: poll unsolditems → age-markdown → RelistWithEdits
Feedback: on SaleCompleted → auto-place positive; alert on negative
Dead-stock mode: Band-C items → $1 no-reserve auction batch

Every box maps to endpoints that exist today. Nothing here is speculative.

6 How automated can it really be? (the honest read)

Workflow	Automation ceiling	Human still needed for
Listing creation	~90%	Approve the batch; fix the odd mangled title/photo
Photo prep	~100%	Reshoot genuinely bad source photos
Pricing	~85%	Set margin floors + markdown curve; override rare high-value items
Buyer Q&A	~70% auto-answered	Judgment-call questions (returns, "will you do $X")
Fulfilment / dispatch	~95% (data side)	Physically pick, pack, lodge the parcel
Feedback	~100%	Service recovery on a negative
Dead-stock $1 auctions	~90%	Decide what's Band C; bundle lots

The honest constraint isn't the API — it's the physical world. Software can draft, price, publish, answer and mark-shipped near-autonomously. A human still photographs the unit and tapes the box. "AI-first" here means the bench operator + the bot do the work of a 4-person listing/admin team — transformative against a business doing all this by hand.

7 Phased build plan

Phase	What ships	Effort	Why first
0 — Sandbox + auth	OAuth 1.0a signing, sandbox app, one listing end-to-end	~2–3 days	De-risks the whole thing; proves the loop
1 — Enrich + Approve + Publish	The photo/IMEI → draft → batch-approve → `POST v1/Selling` pipeline. The crown jewel.	~2–3 wks	80% of the value. Fixes every CRO gap in the store audit
2 — Operate workers	Q&A bot, fulfilment poller, feedback automation	~1–2 wks	Recovers the daily grind; protects seller metrics
3 — Reprice + relist + dead-stock	Age-markdown engine, `RelistWithEdits`, the $1-auction batch tool	~1 wk	Plugs into the bootstrap stock plan's liquidation flywheel
4 — Intel + optimise	Sponsored Listings, competitor-ask monitoring, A/B title tests	ongoing	Margin optimisation once the machine runs

Phase 1 alone — running against the live store's existing 277-listings/day flow — is the step-change. Everything after is leverage.

8 Why this is the right move for this deal

It fixes every CRO gap in the store audit. Titles rarely say "Refurbished", 19/20 listings had zero attributes, 11/20 a single photo, wasted subtitles, no Sponsored Listings, no free shipping — every one is a templatable AI default. The pipeline makes them impossible to get wrong.
It powers the bootstrap dead-stock flywheel. A batch-listing tool turns "4 storage units of aging stock" into $1-auction lots at machine speed — the early auction cash (~$15–30k) that part-funds the warehouse move.
It's the moat the IM never priced. Bolt an automation layer no other TM seller has onto NZ's strongest refurb reputation (19yr, 99.16%, Top Seller) → genuinely hard to replicate. The "activation, not turnaround" thesis.
It compounds your operator advantage. A systems moat alongside the demand-side levers (paid, SEO, CRM) — same operator, more leverage, no extra headcount.

9 Risks & guardrails

T&Cs on data use (real constraint). Trade Me's API terms forbid caching/redistributing their data or building a product whose value is reselling TM market data (Rules 31–35). Keep competitor-pricing intel internal and transient. A private repricing input = fine; a public "TM price tracker" = breach.

Rate limit. 1,000/user/hr is ample but design event-light polling (don't hammer solditems). Request a higher limit for production if needed.
In-trade-only + non-transferable registration (2026). A deal condition + day-one re-registration, not a blocker (§2).
No native auto-relist. You orchestrate relisting yourself — which is better (reprice on every relist).
Reputation is the crown jewel. A mis-answer or mis-ship erodes the 99.16% that makes the account valuable. Keep the human-in-the-loop gate on listings and the judgment-call queue on Q&A until the bot earns trust.

10 Next steps

DD condition

Confirm in writing the in-trade account + feedback history transfer — the single most valuable asset.

Register a sandbox app

Auto-approved. Build the Phase-0 auth loop and prove the listing flow before settlement, zero deal risk.

Decide: direct vs Tradevine

Lean direct on the API.

Spec Phase 1

The photo/IMEI → listing pipeline — the 80/20, doubling as the dead-stock liquidation engine.

Budget it in

A few weeks of build replaces a multi-person listing team — the operational half of the "AI-first" thesis.

Frame it as a timeboxed build, with a fallback. This pipeline is not yet built — it's the Phase 1 project above (~2–3 weeks), run by the same operator doing marketing. Plan it that way: a fixed timebox, and if it slips, a fallback to hire a lister to carry the 277-listings/day workload manually until the pipeline ships. The lean two-operator model is what the business converges to once this is live — not a day-one given.

Grounded in the published Trade Me developer reference and June-2026 eBay automation research. Endpoint paths, rate limits and the April-2026 in-trade registration rule should be re-verified against the live developer docs before committing engineering time. Not legal advice — confirm API T&C compliance for any competitor-intel feature.

→ See the full operating-system vision (Stock Hub)
← Back to the document hub