Lumière PayCheck
Check an x402 endpoint before your agent pays it.
AI agents now pay for APIs on their own with x402: an endpoint answers 402 Payment Required with a price, the agent pays in USDC, and gets the data. Nothing in that flow tells the agent whether the endpoint works, whether the price is right, or whether the payout wallet is the real one.
Lumière PayCheck answers those questions. It monitors every endpoint listed in the x402 Bazaar, grades each one, and gives your agent a plain verdict: proceed, caution, or avoid.
🌐 Live: https://lumierepaycheck.org · 🔌 MCP: https://lumierepaycheck.org/mcp?plans=1 · 📜 Terms
This repository is the public home for docs, examples, and feedback. The monitoring service itself is hosted and closed-source; the scoring formula is public (below).
What it does
-
Monitors every endpoint in the x402 Bazaar (the badge above shows the live count, which grows as new endpoints are listed) every 30 minutes: price quote, payout wallet, uptime, response time.
-
Flags hijack risk, without punishing normal rotations. Every payout-wallet change is checked against the seller's own wallet declaration, the seller's earlier wallets, and direct transfers between the old and new wallet. Confirmed rotations don't affect the grade; unexplained changes are
avoid, thencautionwith human review. Sellers that use a new address per request are recognized automatically. How it works. -
Spending rules for agents. Before paying, an agent asks "may I pay this endpoint this amount to this wallet?" and gets allow or deny with reasons.
-
Plans for teams. Scoped agent keys, daily/monthly spend limits, signed receipts your wallet verifies before paying, a replayable audit trail, and human review for anomalies. Details.
-
Identity and access for agents that pay. Like Entra ID and Intune, for AI agents:
No accounts, no API keys. Free checks are free; paid features are paid per call with x402, the same way agents pay everything else.
Quick start
1. From Claude, Cursor, or any MCP client
Add the remote MCP server:
https://lumierepaycheck.org/mcp?plans=1
- Claude: Settings → Connectors → Add custom connector → paste the URL.
- Cursor / others (
mcp.json):{ "mcpServers": { "lumiere-paycheck": { "url": "https://lumierepaycheck.org/mcp?plans=1" } } }
Tools: check_payment, check_endpoint, report_outcome, top_endpoints, catalog_stats, get_full_report.
Prefer a local stdio server? packages/mcp-server runs the same tools on your machine (node server.js, or the repo's Dockerfile).
2. Drop-in for the x402 fetch client (npm)
npm i x402-paycheck @x402/fetch
import { x402Client, wrapFetchWithPayment } from "@x402/fetch";
import { withPayCheck } from "x402-paycheck";
const fetchWithPay = wrapFetchWithPayment(fetch, withPayCheck(new x402Client().register(/* your scheme */)));
Every payment is checked by PayCheck before it's signed; a "no" aborts it with the reason. Details: packages/x402-paycheck.
Then add one rule to your agent's instructions:
Before paying any x402 endpoint, call the Lumière PayCheck
check_paymenttool with the endpoint URL, the amount, and the payTo wallet from its 402 quote. Only pay ifallowis true. If it returnsallow: false, tell me the reasons instead of paying. After paying, callreport_outcomewith the receipt (or the url and payment tx) and short answers about whether you got what you paid for.
3. From code
// Before paying any x402 endpoint, ask Lumière PayCheck.
const res = await fetch("https://lumierepaycheck.org/v1/check-payment", {
method: "POST",
headers: { "content-type": "application/json" },
body: JSON.stringify({ url: target, amount: quote.amount, payTo: quote.payTo }),
});
const decision = await res.json();
if (!decision.allow) throw new Error(`Not paying ${target}: ${decision.reasons.join("; ")}`);
// ...then pay with your x402 client as usual
More in examples/: TypeScript, Python, curl, webhook and receipt verification, a guarded payer, and plan purchase.
API
Free routes allow 60 requests per minute per client, or 300 with a free API key sent in the x-paycheck-key header. The MCP server allows 300 tool calls per minute. Paid routes use x402 with USDC on Base or Solana (same price on both; on Solana the facilitator pays the fee). Lookups for endpoints we don't monitor return 404 and are never charged. Full reference: docs/api.md.
Plans for teams running agents
Free checks stay free. Plans add authorization for agents that spend money:
Subscribe on the website with a card at lumierepaycheck.org/subscribe: billed monthly by Stripe, cancel anytime, and your account is set up automatically. Then manage everything (agent keys, limits, approvals, billing) at lumierepaycheck.org/account. Developers can also pay with USDC via x402 (prepaid 30 days, no auto-renewal).
The agent calls POST /v1/authorize before every payment and gets allow, deny, or review, with reasons. On allow it gets an Ed25519-signed receipt bound to that exact payment, and examples/guarded-pay.ts shows a payer that refuses to sign without one. Every decision can be replayed later to prove why it was made.
Want it set up for you? The setup service ($750 one-time) configures your agents, limits and approval rules, runs them for a 7-day watch-only week so nothing is blocked by surprise, then hands them over with a guide for your team. An optional monthly review is $250/month.
Full guide: docs/subscriptions.md · Subscribe: lumierepaycheck.org/subscribe · Pay with USDC: examples/subscribe.ts
Verdicts
How a score is made
Deterministic and public. No one can pay for a better grade.
What each grade means: A we paid and it delivered, cleanly and reliably · B we paid and it delivered, with a caveat (reshaped vs its listing, a value warning, 80–95% uptime) · C probably works but unconfirmed (not paid-tested yet, needs its own sign-in, 50–80% uptime, an unconfirmed wallet change) · D doesn't work as listed, but you won't lose money (an error after payment with no charge, or the listing's own example input gets "not found") · F costs money or unsafe (took the money and failed, a payout-wallet incident, under 50% uptime). A/B = proceed, C/D = caution, F = avoid.
- Uptime 40: how often the endpoint returns a valid payment quote over 7 days. Uptime sets the ceiling: an A needs at least 95% uptime and a passed paid test; under 95% the best grade is B, under 80% C, and under 50% F (avoid). These apply once there are 6 hours of checks
- Latency 15: full points at ≤ 1 s median, zero at ≥ 5 s (checks run from one region, so distance isn't penalized). Checks that overlap a stall on our side don't count
- Stability 25: drops with payout-wallet changes and price increases (each distinct higher price counts once, 8 points at most)
- Delivery 20: share of paid test payments that returned what was advertised. Any 2xx counts as success; a 429, a firewall block or an async 202 after payment is neutral; empty result lists count as honest "no results"; a response fails only if it's an error status, empty when output is promised, not JSON when JSON is promised, or has none of the listing's fields (fields wrapped in
data/result, or a list of the listed items, count). Endpoints failing a paid check are re-tested daily - Not yet paid-tested → scored on the other 80 points, rescaled, labeled so, and C at best: A and B mean a real paid test delivered
- Caps: unexplained wallet change → F (unconfirmed after 24 hours → max 70, C); took the money and failed a paid test → F; error after payment with no charge, or the listing's own example input "not found" → D; needs its own sign-in → C; delivers but reshaped vs its own listing (fields nested elsewhere, a list where it shows one item, most example fields missing) → max 85 (B), with a note telling agents where to read the fields; minor gaps (an optional field, a content-type label) get a note only. Confirmed rotations and per-request addresses don't count
- Verdicts:
proceedmeans a real paid test delivered and the endpoint is reliable. Healthy endpoints not yet paid-tested arecaution(delivery not yet confirmed), kept current by our free checks every 30 minutes - Value checks: on each paid test we also check the answer is plausible: it's about what we asked (symbol, address, URL), its "generated at" time is recent, and a major token's USD price is within 10% of the public spot price. A failed value check is a warning with a B cap, never avoid
- A failed payment caused on our side never counts against a seller
- Fair to sellers: only real problems count. Checks where our monitor is rate-limited or blocked by a firewall (Cloudflare, Vercel, AWS WAF, Akamai, DataDome, Imperva, Sucuri) are
blocked; requests the endpoint rejects before quoting (400/405/415/422) aremismatch; failures caused by our own network are . Quotes on payment networks we can't read yet (e.g. ) are . None of these count. Network errors and 502/503/504 are retried once. We send at most 2 requests at a time and about 1 per second to any one host, and pause a host that asks us to slow down
Pricing
For sellers
Every monitored endpoint has a public page and a badge that updates on its own:
[](https://lumierepaycheck.org/e?url=YOUR_ENDPOINT_URL_ENCODED)
Declare your payout wallets so a rotation is never mistaken for a hijack, and a hijack is caught on the first check: publish {"payTo": ["0xYourWallet"]} at https://<your-host>/.well-known/paycheck.json (or a DNS TXT record at _paycheck.<your-host>: payto=0xYourWallet), then POST /v1/declaration with your endpoint URL. Details.
Think a grade is wrong? Open a Grade dispute with the endpoint URL. A bot replies within a minute with the endpoint's current score and failure log, and queues a paid re-test automatically. When the re-test has run, the bot posts the result (grade before and after) and closes the issue if the grade is fixed. Free re-tests cover endpoints priced at $0.05 or less; for pricier ones the bot posts a link to fund a test at the endpoint's own price. Rules and limits: seller guide.
Lumière Agents
From the team behind PayCheck: done-for-you follow-up for small businesses. agents.lumierepaycheck.org
Most small businesses aren't short on work; they're short on follow-up. Lumière Agents does the chasing, in the owner's name and voice:
How it stays safe
- You approve every email at first. Each agent has an "Approve before sending" / "Send automatically" switch the owner can flip any time.
- Facts come from your books, not the AI. Amounts, invoice numbers and dates are inserted by code; the AI only writes the wording, and a draft with made-up numbers or pressure is replaced by a plain template.
- Polite by design. Business hours only, no US holidays, weekly limits per contact, STOP and unsubscribe always work, and nothing is sent to someone who asked to stop.
- Your accounts stay yours. Works with QuickBooks (read-only) or any spreadsheet; emails come from your own address and texts from your own Twilio number. Each client's data is walled off in the database and stored files are encrypted. Pricing: one agent $200/month + $750 setup; each extra agent $100/month + $400 setup; all five $500/month + $2,000 setup. Cancel any time. Guides · Get started
Lumière Agents is a separate, hosted product from Lumière LLC; its code isn't in this repository.
Independence
Lumière PayCheck is an independent project operated by Lumière LLC (Connecticut, USA). It is not affiliated with Coinbase, the x402 Foundation, the Bazaar, or any seller. Scores are automated assessments, not guarantees, endorsements, or financial advice.
Feedback
Questions, feature requests, and grade disputes: open an issue and pick the matching form. Building an agent that pays with x402? I'd love to hear what it needs.
Documentation
Guides for evaluating and running Lumière PayCheck: overview, features, pricing, getting started, administrator guide, integration guide, and FAQ. PDFs: Enterprise plan guide (everything included) and 4-page overview.
Security and privacy
What personal information we collect and who else handles it: PRIVACY.md (also at lumierepaycheck.org/privacy). Live usage numbers: lumierepaycheck.org/stats. Service status and uptime: lumierepaycheck.org/status.
How keys, payments, and data are protected, including current limitations: SECURITY.md (also at lumierepaycheck.org/security). Report vulnerabilities privately via GitHub security advisories or [email protected].
License
The documentation, example code and packages in this repository are MIT licensed. The Lumière PayCheck hosted service and its source code are not part of this repository and are not covered by this license.
"Lumière PayCheck", "Lumière" and the Lumière PayCheck logo are trademarks of Lumière LLC. The MIT license covers the code, not the names or logo: forks are welcome, but please give them a different name.