Kelp Market · For AI labs · 2026
Train on data you
can defend.
Kelp datasets hold images no model has ever trained on — most were never published anywhere online. Every asset ships with per-asset proof: where it came from, who licensed it, which camera took it. Your agents quote and buy over MCP. The paperwork is already done.
ed25519 · SEAL records · PRNU sensor fingerprinting · Stripe rails
Why now
Documentation decides these cases
Getty lost most of its UK case against Stability — but the judgment turned on what Stability couldn’t document about its own training runs. Paperwork decides these cases.
EU AI Act Article 50 went into force. Training on data you can’t document now has a compliance price attached.
Stability is in front of a US jury. However the verdicts land, per-asset provenance becomes either a defense exhibit or a settlement ledger.
The inventory
Assets no model has ever seen
~13.6 trillion unique photos sit in camera rolls, drives, and archives — Rise Above Research, 2026.
~136 billion are indexed on Google Images. That is the supply every lab has already trained on — scraped, recycled, in every model.
80%+ of the world's photos were never posted anywhere online. That is Kelp's supply: partners' unpublished collections, verified at the gear, licensed as private datasets. Never posted, never resold as stock.
What you get
Every claim comes with evidence
01 — Signed manifest per asset
Asset id, bundle location, license, rights holder, price — each certificate signed. Altering a single byte is detectable.
02 — Camera-verified provenance
Three proof tiers: EXIF self-report, gear challenge, or PRNU sensor fingerprint. Priced by proof strength, so you choose.
03 — Fail-closed verification
One verifier, one public key published in DNS. If any check fails, no verification badge. No trust lists, no validator disagreements.
04 — MCP-native buying
Your agent estimates, pays, and downloads over the Model Context Protocol. No sales calls, no data-room PDFs.
How buying works
Quote to bundle in three calls
01 — Estimate
Call estimate_training_set with a type, quantity, and keyword. A real quote comes back in seconds — no key needed.
02 — Confirm and pay
confirm_training_set locks the quote and queues delivery. Settled via Stripe Checkout; the quote expires in an hour, so no stale pricing.
03 — Download the bundle
A tarball with LICENSE.md, manifest.json, and a signed proof per asset. Everything a court or auditor asks for is already in the file.
Pricing
Proof strength sets the price
EXIF self-reported
$0.50/asset
the words in the file say what it is.
Challenge verified
$1.00/asset
the gear answered our challenge — camera confirmed.
Sensor fingerprint
$2.00/asset
PRNU binds the image to that physical sensor.
volume discounts: 10% at 100 assets · 20% at 1,000 · 30% at 10,000 — combined with coupons, capped at 40%
For AI labs
Start with an estimate.
list_datasets and estimate_training_set need no key