Buy it. Sell it. See whether it actually works.
Models, agents, tools, datasets and whole AI systems — from a nine-euro logo to a fifty-thousand-euro system. Nothing reaches the shelf without a benchmark we ran — and when the benchmark is not done yet, the sheet sits there, empty.
The problem, plainly
Someone builds a local music AI at home that beats what is on the market, and wants fifty thousand euros for it. Where do they put it?
- Not in a vendor's own store. You sell on Salesforce only if it runs on Salesforce. In the GPT Store only if it is a GPT. Every AI marketplace is somebody's shelf, and half of that shelf is already theirs.
- Not in a file repository. You can upload the weights, but there is no price and no buyer there.
- Not with the business brokers. They want companies with recurring revenue, not software.
- And not on a classifieds site. Between a sofa and a bicycle.
Today the answer is: nowhere. That gap is the market, and that is why we exist.
Who sells and who buys
These are not four marketplaces. It is one grid with two axes. We open where we already have goods, and we say plainly where there is nothing yet.
Work done by machines
Images, text, product copy, translations, manuals. Produced automatically, bought by an ordinary person with a card.
Tools for AI agents
API endpoints, tools, datasets, modules. This is where we already have goods that work and have a paying client — the manual tool. The endpoint an agent calls directly opens together with the catalogue.
Machine-to-machine trade
The standard exists, but the room is practically empty: real measured agent-to-agent commerce is under fifteen million dollars across the entire life of the protocol, and falling. The room stays open, and costs nothing to keep.
Human to human
Full, and it has owners: Gumroad, Fiverr, Etsy. We are not opening a market here — but what we know is sold here too.
Why buy here
We do not just list. We measure.
Nobody pays fifty thousand euros on the word of a stranger on the internet. And no existing marketplace solves this — the listing says „translation, two dollars" and nothing else. Our rule: nothing reaches the shelf without a standard benchmark we run, and the result sits next to the price with the method written down. Where the benchmark is not done yet, the sheet is empty and visibly so.
And this is what the big ones cannot do: a referee who is also a player is not a referee. Salesforce and Microsoft may not benchmark other people's goods, because half the shelf is theirs. We have no shelf of our own to protect.
Fixed-input benchmark
Same inputs for everyone. We publish times and outputs.
A sample we ran
We push a test file through it and show what came out.
Counted, not described
How many rows, what format, how many gaps, where it came from.
A recording of a run
Plus the measurement sheets, exactly as they came out.
House rules
Every number on this site carries its label — and the shape of the number tells you what it is before you read it:
- MEASURED a single number It happened. On a specific machine, on a specific date, and both are written next to it.
- ESTIMATED a range Computed from a measurement. The formula is written down, not hidden.
- NOT REAL YET no number at all A hope, an offer, a „they said they would". We give it no number, because it has none.
A number you cannot verify is advertising. A number with its method written beside it is a claim.