Get the tool list and price sheet from MCP Registry, llms.txt, or the marketplace catalog.
Contains the four elements network / asset / amount / payTo; the client learns the price without registering.
The buyer only needs to hold USDC; gas is sponsored by the facilitator. After signing, replay the request with the credential.
No charge when a failure returns isError; on success it goes on-chain, and the platform collects then splits the revenue.
Clients complete the first four steps in the console; the platform handles endpoint generation, settlement integration, and discovery-layer registration.
HTTP API (with auth and polling), CSV / file upload, direct PostgreSQL connection, scheduled crawlers. After connecting, you can preview retrieval results directly and confirm data freshness before slicing into tools.
Slice the data source into the smallest paid units that return within 1 second, then define name, description, input parameters, return fields, and price one by one. The platform checks whether the description is Agent-readable — tools written unclearly will not be bought.
At least one free tool (showcase) is required; paid tools default to $0.05. A unit price below $0.01 makes no sense in this system — each facilitator transaction costs about $0.001, so pricing too low means losing money on every sale.
FastMCP exposes the tools + x402 payment middleware issues the 402 challenge + Base mainnet + Circle USDC + PayAI facilitator. Every link in this chain has pitfalls; we hit them all, so you only need to click Generate endpoint once.
No settlement when a failure returns isError — failed delivery is not charged; that is the foundation of trust.
The official x402.org facilitator only supports testnets, and Mogami only has v1 — the cost of choosing wrong is an entire chain that does not work.
Version upgrades and facilitator switchover drills — for teams that do not want to run their own ops.
tools: - name: ping price: "0" # 免费层 desc: 健康检查 - name: list_gpu_models price: "0" # 橱窗 desc: 全部 GPU 卡型清单 - name: gpu_price price: "0.05" params: { gpu: string } desc: 指定卡型的当前最低价、中位价与 30 天最低价 - name: gpu_history price: "0.05" params: { gpu: string, days: int } desc: 30 天价格历史、最低价平台与库存状态
We call it GEO (generative engine optimization). It is not mysticism — it is a five-rung ladder you can work through rung by rung.
Explicitly allow ClaudeBot, GPTBot, OAI-SearchBot, PerplexityBot, Google-Extended, and add PetalBot, Yeti, Daum, Qwantify and a dozen more sources based on measured logs.Block them, and the next four layers are all for nothing.
Write the JSON-LD Dataset node clearly: the downloadable dataset, how many retrieval entry points, update frequency, covered time range, and which fields it contains. Let AI know at the machine level where to fetch data.
Put a complete answer at the top of the page that can be repeated verbatim: who, how many, when, source URL, license. Large models tend to cite well-formed sentences rather than assemble them from tables themselves.
In /llms.txt, state what the site is, how the data is used, how much is free, and how to cite it (How to cite). This file is becoming the de facto standard for AI crawling.
The layer most people miss: after AI mentions you, users who want to dig deeper need a path that requires no registration and no waiting for a sales reply.Citations bring attention; settlement turns attention into revenue.
Submit to the official MCP Registry (publish the tool as a binary via GitHub releases — the package with the same name on npm is fake), llms.txt directories, and AI-friendly site indexes.
Featured slots on the data marketplace homepage, category pinning, on-site cross-recommendations. Clients under the collection model can apply to join joint promotions.
Crawl logs split by crawler source, AI session referral traffic stats, direct-answer card hit counts. You can only fix a layer you can see.
The platform collects USDC → aggregated by calendar month, settled on the 10th of the next month → settlement statements are exportable. Reconciliation trusts only Transfer events.
eip155:8453
Official Circle contract
Gas-free sponsored authorization
No balance deltas
Under the collection model, each transaction amount is distributed in this order:
The same H100 has a dozen different quoting conventions across a dozen platforms. Without unified fields, Agent cannot make decisions for humans.
| Field | Required | Definition and constraints |
|---|---|---|
| platform_id | Required | a lowercase stable identifier.Does not change when the brand is renamed; historical continuity takes priority over readability |
| gpu_model | Required | model + VRAM, e.g. H100 80GB; no abbreviations, no aliases |
| unit | Required | fixed to hour, all other units are converted by the consumer and never mixed into the source data |
| currency | Required | ISO 4217 (CNY / USD), with the exchange rate reference time |
| billing | Required | on-demand / reserved / spot One of three,Different billing types are not allowed in the same ranking |
| stock | Required | in-stock / limited / out; Out-of-stock quotes must be explicitly marked |
| region | Required | Platform registration / datacenter region (China / North America / Europe / Asia-Pacific…) |
| includes | Optional | List of items included in the price (CPU / memory / storage / network); missing means 'not declared', not 'all included' |
| observed_at | Required | ISO 8601 fetch timestamp; together with refresh_minutes it expresses freshness |
| license | Required | Citation license, CC BY 4.0 recommended; also specify the attribution format |
Then take a share from the increment. Compute platforms can also offset service fees with existing referral programs or per-click settlement.