What Agents Buy

com.whatagentsbuy由 whatagentsbuy.com 發布官方發布者v1.0.1更新於 Oct 24, 2026

Independent measured reviews of x402 API sellers: pre-payment verdicts and accuracy grades.

已驗證Streamable HTTP可網頁執行金融安全與監控資料與分析

概覽

AI 產生的概覽

一個遠端 MCP 伺服器，為助理提供對 x402 API 賣家的獨立實測評價，包含付款前安全判定與準確度評級。

可以做什麼
依任務描述尋找可付費的 x402 API，並依實測準確度、付款安全性與價格排序，同時查詢特定賣家的完整紀錄、報價與實際扣款及需求評分。在付款前對某個 URL 或主機給出 CLEAR、HOLD 或 ABORT 單一判定，並可取得詳細付款安全解讀。依收入或真實需求對市場上的賣家排名，並依回傳數值與一手來源的接近程度對客觀類別排名。查看 x402 市場規模、Base 近期結算與 24 小時交易量趨勢。閱讀已知付款陷阱的實地紀錄與規避方法。
適用情境
適用於助理即將向某個 x402 API 付款、需要安全檢查時，在多个可付費 API 之間為任務做選擇時，或評估某賣家資料可信度時。也適合回答 x402 市場規模與結算活動相關問題。
執行需求
遠端 streamable HTTP 端點；未宣告需要身分驗證、付款、環境變數或請求標頭。需要能存取該端點的網路。
安裝前請注意
伺服器本身不收款，但其判定與排名用於把關在別處進行的真實 USDC 付款，因此錯誤或不確定的判定可能造成資金損失。相關結論自稱獨立實測且無贊助，但屬於第三方評估；遇到 HOLD 或無法測定的結果，應在付款前自行核實。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 What Agents Buy，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Web executable，透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。

交給代理程式安裝

把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式，或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南，先向你展示發布者、端點和登入方式，經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。

閱讀 https://sourceweft.com/mcp/com.whatagentsbuy%2Fwhatagentsbuy/install.md，並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前，先告訴我它的發布者、端點和登入方式，等我確認。不要把 API 金鑰寫進命令或我會提交的檔案；問我金鑰應從哪裡取得。

檢視代理程式所依據的安裝指南

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "what-agents-buy": {
      "type": "http",
      "url": "https://whatagentsbuy.com/mcp"
    }
  }
}

README

What Agents Buy

Independent, measured reviews of the x402 APIs that AI agents pay for.

Live at whatagentsbuy.com · MCP server at https://whatagentsbuy.com/mcp · listed in the official MCP Registry as com.whatagentsbuy/whatagentsbuy.

Most numbers quoted about agentic payments come from registries and press releases, and they are wrong by orders of magnitude. This project measures the market directly. It pays x402 endpoints with a real wallet and publishes the receipt (what was quoted, what was actually charged, whether the goods arrived), sweeps every USDC payment agents make on Base straight off the chain, and grades the number a seller returns against a primary source that cannot be a reseller. Nothing is sponsored. No service pays to appear or to be graded.

This repository is the open methodology: the measurement pipeline and the agent-facing MCP server. The accumulated settlement history and the archive of paid responses stay private; everything they produce is published on the site.

What it measures

  • Settlement. USDC Transfer logs on Base, addressed to the payment wallet each service declares in its own x402 challenge. Measured from the chain, not from a registry's self-reported counters, which have been wrong by 30x. (sweep.py, leaderboard.py)
  • Delivery. Every graded call is a paid call. A field counts as delivered only if it is actually present in the response, checked against the seller's own schema, and every short verdict is re-verified with a second independent call before it is recorded. (conform.py, receipts.py)
  • Accuracy. For categories with an objective ground truth (crypto price, stock quote, FX rate, gas, wallet balance, weather), the number a seller returns is graded against a primary source that cannot be a reseller: an exchange median, a professional feed, the ECB, the chain itself. (lab.py, categorize.py)
  • Preflight. One CLEAR / HOLD / ABORT verdict per seller, folding live price and payTo honesty versus the listing, phantom paywalls, and delivery receipts into a single signal an agent can gate a payment on. A red light fires only from hard evidence. (preflight.py)

Every harsh finding is verified before it is published. Several early accusations turned out to be measurement bugs in this code, so the pipeline runs known-good canaries before it grades anyone, treats "we could not measure it" as inconclusive rather than a negative verdict, and re-verifies every shortfall twice.

The MCP server

api/mcp.js is a stateless, dependency-free MCP server (streamable HTTP, no auth, no payment) that exposes the measurements as tools an agent calls before it pays:

  • preflight(url) — the one pre-payment verdict; detail:true adds the full payment-safety read.
  • find_api(task) — discover a payable API for a task, proven-accurate sellers ranked first.
  • most_accurate(category) — rank an objective category by how close each seller's returned value is to a primary source.
  • get_service, search_services, top_services, list_traps, market_summary.

Add it with:

claude mcp add --transport http whatagentsbuy https://whatagentsbuy.com/mcp

Pipeline

fetch_registry.py    # discover the payable endpointssweep.py             # Base USDC settlement tape (sweep_solana.py for Solana)whatsnew.py          # registry diff, first-seen trackingprobe.py             # free directory probe: who answers, price/payTo honestyconform.py           # the paid-call chokepoint: archive raw, judge, reconcilelab.py               # daily accuracy corpus vs primary sourcesreceipts.py          # content-hashed, self-verifying dispute receiptsleaderboard.py       # settlement to a leaderboard + Organic Demand Scorepreflight.py         # the pre-payment verdict enginebuild.py             # everything to a static site

The *_shootout.py scripts buy the same thing from every seller in a category (email validation, gas, WHOIS, price, weather) and publish the contrast.

Tests

python3 tests.py            # pipeline invariants, fixture-based, no networknode --test api/*.test.js   # the MCP server

Both run in CI on every push, and tests.py runs at three gates locally: before every build, in the pre-commit hook, and before any script spends money.

Author

Neil K. Patel — neilkpatel.com · LinkedIn

Feedback and corrections are welcome on LinkedIn.

License

The measurements published on the site are CC BY 4.0 (attribute What Agents Buy). This code is shared for transparency and independent review of the methodology.

來源：README.md，提交 39b1b61

工具

7

find_api

Find a payable API for a task

USE WHEN you need an API that does something and want payable options. Searches the whole x402 market for what you describe and returns a ranked, pre-vetted shortlist, proven-accurate sellers first, then payment-safety, then price. Covers every reachable endpoint, not just ones we have bought from. Next step: check_before_paying on your pick, then read the payTo from the live 402 and sign against it.

輸入參數task必填limitmax_price_usdcmin_reliabilityallowed_verdicts

check_before_paying

The one check before your agent pays

USE WHEN your agent is about to pay an x402 API and you want to know if it is safe and worth it. Returns one verdict for a URL or host in the `verdict` field (gate on that; `light` is the same decision as a colour, for display): CLEAR (nothing alarming), HOLD (pay but resolve the reasons), ABORT (do not pay without checking the live 402), UNRATED (no data). Also returns a confidence tier (verified = a payment to this seller settled AND produced a gradeable result, checked = free probe measured) and its payment history. Folds live price and payTo honesty, phantom paywalls, delivery receipts, and the wash/real-demand read into that one light. Pass detail:true for the full read. Whatever it says, still read the payTo and amount out of the live 402 and sign against those. (Formerly named preflight.)

輸入參數url必填detail

look_up_seller

Everything known about one seller

USE WHEN you have a specific seller host and want its full track record. Returns every grade earned by actually paying it, what was quoted versus charged, whether goods arrived, its settlement volume, and its Organic Demand Score with the demand-shape read (is the volume real independent demand, or one wallet supplying almost all of it). The deep dossier on one seller; for a quick go/no-go before paying, use check_before_paying instead. (Formerly named get_service.)

輸入參數host必填

rank_sellers

The leaderboard: biggest, or most real

USE WHEN you want a ranked list of sellers across the market. by='revenue' ranks by raw USDC received (who is busy); by='real_demand' ranks by Organic Demand Score (whose money comes from many independent wallets rather than one wallet supplying almost all of it). The two disagree often, and the disagreement is the point. For accuracy ranking within an objective category, use rank_by_accuracy. (Formerly named top_services.)

輸入參數bylimit

rank_by_accuracy

Who returns the most accurate value in a category

USE WHEN a category has an objective right answer (crypto-price, stock-price, fx-rate, gas-price, wallet-balance, weather) and you want the sellers we PAID ranked by how close their returned value was to a primary source that cannot be a reseller (exchange median, chain balanceOf, ECB rates, FMP quote). Top rows are the cheapest accurate sellers. Call with no category to list the available ones. (Formerly named most_accurate.)

輸入參數limitcategory

market_size

How big is the x402 market

USE WHEN asked how big x402 is, whether it is growing, or what settled recently. Returns both what actually settled on Base over the last day (from chain logs, per tracked seller) and the live market pulse (24h volume and payments with day-over-day change, a 7-day series, and stablecoin circulation). All measured from chain, not registry counters, which are not demand and can be bought. (Merges the former market_summary and market_pulse.)

無需輸入參數。

known_payment_traps

How agents lose money paying APIs

USE WHEN writing payment code, or before trusting an unfamiliar seller. The field notes: each is a specific way an agent paying an API loses money or is misled, what it cost to learn, and what to do instead. (Formerly named list_traps.)

輸入參數query

版本歷史

1
  1. v1.0.1最新Sep 16, 2026

評分

已滿足適用檢查項的 90%

該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通必需25 分
  • 已宣告認證方式必需10 分
  • 已列出工具必需15 分
  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分

    請在儲存庫或註冊中心條目中宣告授權條款。