agent-receipt-verify

Independent verification of payment- and outcome-proof receipts used by AI agents.

Run it, one line

npx -y agent-receipt-verify https://api.bitrefill.com/x402/gift-cards/search

No account, no key, no payment. It tells you whether that endpoint asks for money the way x402 says, and says NOT MEASURED for the levels that need a real settlement instead of failing them.

What it does NOT do: it judges the SHAPE of a receipt, not the honesty of the party that issued it. See CONFORMANCE.md for what is deliberately left unmeasured.

npx -y agent-receipt-verify attempt-identity   # 72 vectors for x402 spec section 5.5
npx -y agent-receipt-verify link-vectors       # 14 vectors (13 offline), receipt to settlement
npx -y agent-receipt-verify kanon              # cross-check against iamonuwa/kanon

Related MCP server: x402-test-mcp

What this is

An agent economy has two kinds of receipt. One says who was allowed to act (delegation: UCAN, Authproof, Agent Passport). The other says what was paid for and whether the output matches (settlement and outcome: x402 offer-receipt, INAM countersigned receipts, and the proofs ERC-8004 consumers count).

receipt-commons/receipt-interop covers the first kind. This covers the second.

Use it from an agent: the MCP server

Both checks are also a Model Context Protocol server, so an agent can run them itself instead of a person running a command. It speaks MCP over stdio, protocol version 2025-11-25. The framing, initialize, tools/list and tools/call are implemented directly in src/mcp-server.mjs, with no MCP SDK and no required dependency at all.

Two tools, for the two moments where an agent actually needs an answer.

check_x402_endpoint, before paying. Sends one request and reports which levels of the ladder in CONFORMANCE.md the endpoint reaches. Levels that need a real settlement come back as not measured, never as failed.

link_receipt_to_settlement, after paying. Takes the receipt, the authorization salt and the asset that was paid, resolves the salted commitment against live chain state, and answers linked, not linked, or could not check. Those are three different answers and it never collapses them into two.

Add it to a client:

{
  "mcpServers": {
    "agent-receipt-verify": {
      "command": "npx",
      "args": ["-y", "github:holistis/agent-receipt-verify", "mcp"]
    }
  }
}

No account, no API key, nothing published to npm. If you would rather not run Node from a git ref, there is a Dockerfile:

docker build -t agent-receipt-verify . && docker run -i --rm agent-receipt-verify

The server opens no port and writes nothing to disk. Anything that is not a protocol frame goes to stderr, because stdout belongs to the client. npm test drives the server over real frames in 21 cases, including an unknown tool name, a notification that must not get a reply, and a network it does not know, which has to answer could not check rather than not linked.

ethers is an optional dependency, used for one signature check inside check_x402_endpoint. Without it that single check reports not measured instead of passing, and everything else works unchanged.

Why independent implementations matter

A receipt signed by a seller is a statement by an interested party. It only becomes evidence when someone who was not the signer can recompute it and reach the same conclusion. So every verifier here is written against the published specification, never against the reference implementation's source, and is tested against the vector files published by the protocol's own authors.

If two implementations that have not seen each other's code agree on the same vectors, that is interop. One implementation agreeing with itself proves nothing.

Who has independently taken this up

These are other people stating something about this project, not claims made here. Every line is a link you can open.

• INAM names it in their own verifier documentation as an independent implementation, written from SPEC.md and not from their source, and adds that they have not audited the code and that you should run the vectors yourself: docs/VERIFIER.md

• goun7 carries two measurements from this project in the honest-limits section of his own vector document, with the wording kept as ours: TEST-VEKTOR-KOK-COMPUTE-WORK.md

• iamonuwa decided to link this project's receipt-to-settlement corpus from Kanon as a complementary set with no endorsement, and to keep the two formats compatible: kanon#32

• babyblueviper1 deployed a refusal reported here, where a producer carrying a display name and no key id is rejected before payment settles: erc-8004#99

Two of these carry a caveat worth repeating. The Kanon README link is stated in the issue and is not live yet, and INAM says plainly that they have not audited this code. Both are written that way on purpose. A list of people who looked is worth more when it also says what they did not do.

Status

npm test      # all of the above. No account, no key, and nobody else's server

Node built-ins only, apart from keccak256. Two honest footnotes, because "no network" was not quite true and is the kind of claim this package should not make loosely:

• One case dials a closed port on localhost on purpose, to prove that a dead address reports not-measured instead of crashing the tool. That is a socket, but it leaves the machine alone and reaches no third party.

• Until recently the suite also fetched a live discovery index belonging to someone else, roughly twenty thousand entries, because one test handed an unknown flag to census and census ignored unknown flags and went to work. That flag is now refused, so the suite finishes in about eleven seconds and contacts nobody. A single index round alone used to take over thirty.

Only the settlement link below reaches a real network, and then only a public, read-only RPC.

Confirmed from the other side, which matters more than the table: the author of the x402 specification ran their own implementation, which shares no code with this one, against all three live receipts below and reached the same three settlements, starting from the receipt, the salt and the asset as §5.6.4 requires. Their result and ours are in x402#3827.

Recompute it yourself, without an account

Two of the cross-checks run offline from a single command, so nothing here has to be taken on trust:

npx -y agent-receipt-verify kanon
npx -y agent-receipt-verify beleid
npx -y agent-receipt-verify link-vectors

The first checks our EIP-3009 verifier against the ten published vectors of iamonuwa/kanon. The second exists because agreement can hide a question: all ten of those vectors cover underpayment only, so they say nothing about overpayment, or about one nonce reused across two payers. Those are the two open questions in x402#2725, and the command prints exactly where the two possible answers diverge and where they do not.

One scope to know about. The high-s rule and the signer recovery apply to a payer WITHOUT code. USDC picks the check by the payer's code: no code, ecrecover (and a high-s signature reverts); code, ERC-1271 isValidSignature on the payer's contract. So a smart wallet, or an ordinary address with an EIP-7702 delegation, must not be rejected for the shape of its signature. Pass payerHasCode: true in the context and the verifier skips both rules, still applies the window, the amount and the nonce, and never answers VALID: it answers ongemeten, because the signature itself is only checked by a call to the chain. Every answer now carries payerScope (code, no-code or unknown; unknown behaves as no-code and says so). Pointed out on x402#3925 by robertolocatelli81-dev, and it applied to our first version too.

The third runs fourteen vectors for the half AFTER paying: does this receipt belong to that one settlement. Thirteen of them need no network at all; the fourteenth reads the chain and is marked and skipped rather than reported as a failure, because a check that could not run says nothing about the receipt. The same twelve also exist in the field layout of the Kanon vectors, in vectors/settlement-link-kanon, so the two corpora can sit side by side.

Two more commands, both of which measure other people's endpoints rather than our own, and both read-only:

npx -y agent-receipt-verify census [out.json]
npx -y agent-receipt-verify liveness [out.json]

census counts a discovery index per HOST rather than per resource, so a party with three hundred paths does not outweigh three hundred parties with one. liveness takes that census and asks each host whether it still answers 402 today, because a registration in an index is not a promise that the endpoint is alive. Both are slow by design: they pause between requests, because the hosts in that list did not ask to be measured.

Linking a receipt to its settlement

A receipt with an empty transaction field can be matched to a payment only by payer, payee, amount and time. Two payments with the same terms in the same block cannot be told apart at all. Section 5.6 of the x402 offer-receipt extension fixes that with a salted commitment to the ERC-3009 nonce: whoever holds the salt can recompute the link to exactly one settlement, and everyone else learns nothing.

Where section 5.6 lives, because it is not where a reader would assume. Sections 5.6 through 5.6.5 do not exist in the version of the offer-receipt extension on main. They exist only in pull request 3827, which is open. Everything here that cites 5.6 is therefore measured against a proposal, pinned to commit 0e492d6 on that pull request, which is four commits ahead of main and fetchable by anyone. The numbering can still move before it merges. Levels 4 and 5 of the ladder carry this in the profile, and the tool says so in its own output, so a result is never quietly presented as measured against published text.

Section 5.6.5 is the part that makes this matter in practice. A commitment that nobody has opened is still only a statement by the signer: they can issue two receipts with different salts for one settlement, or a commitment that opens to nothing at all. So a consumer that counts settlements, for a deduplicated or Sybil-resistant count, may count only the commitments that someone has opened.

Here are three payments with identical terms: same payer, same payee, same amount, same resource, minutes apart on Base Sepolia. Without a commitment those receipts could not be told apart. Each one still resolves to its own settlement:

npm run link-example

The default window is 800 seconds either side of the receipt's issuedAt, translated to blocks with the pace measured around that spot; --window=<blocks> sets it in blocks instead. If a public node has pruned the block, the command says so, then asks a keyless explorer (Blockscout) for the same logs, and the output names which reader answered (reader :). An explorer is a third party between you and the chain, not a node: it reads the same chain but can behave differently, so the node stays the first choice and the explorer is a second opinion. Use --no-explorer to forbid the fallback, or --explorer=<url> for another Etherscan-compatible API. If both fail you get LINKED : unknown, which is an incomplete check and never a verdict.

Erratum, found 24 October 2026 (versions up to and including 0.2.11). The command placed the receipt's time on a block by assuming the pace at the chain tip held for the whole way back. Base Sepolia made a block every 2 seconds on 16 October and one every 0.2 seconds a week later, so the estimate landed 1.9 million blocks too early. On a node with full history that searched a window where the settlement was not and answered LINKED : no, which is a wrong answer; on public nodes it failed with "pruned history". A stranger ran the command six times and never got a verdict, which is how we found it. 0.2.12 searches for the block instead of computing it, measures the pace around the block it finds, and uses a window in seconds. If an earlier version told you LINKED : no for a receipt that is older than a change in a chain's pace, run it again: that no was about the wrong window.

commitment  : 0x5c35a61a09a566a9c284fad18a70ac2ddfff34d01d2a3ce776ce94e8d5b38ff8
candidates  : 2 nonce(s) consumed by this payer in that window
LINKED      : yes, to exactly one settlement
  tx        : 0xed87b09bfb20a8c8b69908e163fcea6432d193d389640c0465afcb5561dfeb30

commitment  : 0xfecee93e1aec7d6858a218bf324f23e2a4eed07925d8d4542b9662c5026e3d22
candidates  : 3 nonce(s) consumed by this payer in that window
LINKED      : yes, to exactly one settlement
  tx        : 0x2898d0ceb97ddf60af98e19c94fc0adddbdf1e64f9032e62454085e37acfe36b

commitment  : 0xea8dd34b8beac90f021a4e8d893a689c25249f86947770c8e6647c1cde206008
candidates  : 2 nonce(s) consumed by this payer in that window
LINKED      : yes, to exactly one settlement
  tx        : 0x2a9e95d63a5af8210767baa69143b364dad51caffc29a152c77e03bc508d7199

Each window is centred on its own receipt's issuedAt, which is why the candidate counts differ: the middle receipt's window covers all three payments and it still picks out exactly one. That is the ambiguity section 5.6.1 describes, resolved against live state.

Two things in that output move and the verdict does not. The candidate count rises as that payer consumes more nonces, and the upper block number is capped at the current chain head until the head passes it. The transaction each receipt resolves to is fixed.

The four files in examples/ hold exactly what section 5.6.4 says a later verifier has: the receipt, its salt, and the asset that was paid. The nonces and the transaction hashes are deliberately kept out of them. Those come out of AuthorizationUsed(address,bytes32) logs on the token contract, and the tool accepts an answer only when exactly one candidate opens the commitment. Zero matches and more than one match are both reported as not linked, never as a best guess.

Point it at your own receipt with --receipt=<file>, and add --salt= and --asset= if they are not stored alongside it. All three are required: a receipt and a salt without the asset cannot link, which is section 5.6.4's own warning.

Those payments went to a live x402 seller that issues version 2 receipts, and it publishes these cases itself, with the settlement each one should resolve to, at https://api.tokenizen.nl/api/x402/conformance. Point your own implementation at that and compare, or call its paid route and link a receipt you obtained yourself. A fixture cannot tell you whether the salt arrives at the location §5.6.3 names, inside a real settlement envelope; only a paid call can. Buyer and seller are both us, so this shows the chain closes end to end. It does not show that anyone else is using it.

Three of the four are on Base Sepolia. The fourth, receipt-live-base-mainnet.json, is a real payment of 0.001 USDC on Base mainnet on 2 November 2026, with the registration, the payment and the feedback on chain (agent 1150884). Someone outside our team, ruiruii, ran the ground command from a clean folder on the same day and reported that all four facts hold, that the feedbackHash opens on that settlement, and that the binding reads grounded at the payment block, the feedback block and now (erc-8004/erc-8004-contracts, issue 99, comment 6578898622). One line stays not measured: the rule that the payment lands after the agent was registered, because a public RPC will not serve the archive read for the registration block. Pass --rpc= with an endpoint that serves log ranges, and for the registration rule an archive node. Buyer and seller are both us, so this too shows the chain closes, not that anyone else uses it.

Is an ERC-8004 feedback grounded in a payment?

npx -y agent-receipt-verify ground --feedback=0xa661b2ffd344926ddd5e31dc0c481212ba2d973fdb0c08a461d4c903f15d3936

Four facts, read from the chain and reported one by one: the rater is the payer, the paid address is the agent's bound wallet, the payment lands before the feedback, and the rater is neither owner nor approved operator. They are the definition AgenticFinanceGraph uses in erc-8004/erc-8004-contracts#99, rebuilt here from their description and not from their code, the way the Kanon vectors were. A fifth fact stands apart: given the receipt and its salt (--receipt= --salt=), does the feedbackHash open, through the receipt, on exactly that settlement, so that a reader does not have to believe the hash points at the right transfer.

Checked against three verdicts someone else wrote down before this existed, recorded from the real chain in vectors/grounding and replayed offline by test/grounding.mjs and by ground --vectors: agent 1085288, where all four facts hold, agent 524087, where the rater is the payee and the payment did not go to the agent's bound wallet, and agent 10402 on Base Sepolia (case 1 of predge-ai), where the bound address changed between the payment and the feedback, so the payment-block reading holds and the feedback-block reading fails. Payer and payee come from the USDC transfer log and not from the transaction, because a facilitator sends the settlement. The binding is read at the block of the feedback; a node that cannot do that gets a read at the latest block and the output says so. A fact that could not be measured is not_measured and never false. The verdict is about the chain facts, not about whether the rating is fair.

Since 0.2.15 three more facts are shown next to the four, and they never change the verdict. They exist because AgenticFinanceGraph named two conditions in its count that its description had left out (#99, 25 Oct 2026), so that a reader can apply either rule and keep counts comparable:

• paymentAfterRegistration, with registration.blok and paymentBeforeRegistration in the JSON: was the payment made after the rated agent was minted? Before the mint the payee was an address, not the agent. The block is found by bisection on ownerOf and needs a node that serves historical state; otherwise it is not_measured.

• noReturnTransfer: no transfer of about the same amount (at least 90 percent) comes back to the payer in the same transaction, which would make the payment a routing hop.

• binding.agentWalletNu and binding.koppelingVeranderd: the bound wallet at the feedback block and as it stands now, and whether it changed. We read the binding at the feedback block; an indexer that counts later reads the later one.

Only stablecoins identified by the network's USDC contract address are read; other tokens are not.

Since 0.2.16 there is a third reading of the binding, paidIsBoundWalletAtPayment with binding.agentWalletBijBetaling: which wallet was bound to the agent at the block of the payment. It answers whether the payment went to the agent's registered wallet when it was made, and it covers the pre-registration case without a separate rule, because before the mint there is no binding to read. Since 0.2.21 the output ends with the three readings side by side (payment block, feedback block, now) and says so when they differ, because the GROUNDED line above them is the feedback-block reading and can say yes where the payment-block reading says no (case 2 below). The tool does not pick a reading for the reader. Which one a grounding rule should require is a spec question; this only makes the three answers visible.

The three readings have vectors in vectors/grounding-readings/erc8004-grounding-readings.json: three cases with the expected verdict per reading. The same three cases are replayed offline on a simulated chain by test/grounding.mjs. The expected verdicts and the first two cases were written first by predge-ai (erc-8004/erc-8004-contracts#99); the third case, where the binding changes only after the feedback, is ours. All three cases are recorded live on Base Sepolia (case 1 by us, agent 10402; cases 2 and 3 by ruiruii, agents 10419 and 10420) and the file lists their transactions, so anyone can run ground on them. ground --vectors replays five recordings offline, and says which of the expected verdicts were written by someone other than us. scripts/neem-grounding-op.mjs makes a new recording.

One suite

npx -y agent-receipt-verify suite

Four checks in the order of the chain, one table, exit code 1 if any of them fails: the attempt identity of x402 section 5.5, the payment authorization against the published EIP-3009 vectors, the receipt that must open on exactly one settlement, and the feedback that must be grounded in a payment (the two recorded verdicts of AgenticFinanceGraph, replayed offline with ground --vectors). It was asked for in erc-8004/erc-8004-contracts#99, and it adds no second implementation of anything: it calls the commands above and puts their answers side by side. It checks that this implementation behaves like the published vectors and the recorded verdicts of others. It is not a certification of anyone else's endpoint, receipt or feedback.

Honest limits

The format checks test the form of a receipt: signatures, content addressing, and whether the hashes match the bytes a counterparty observed. They do not tell you whether an operator is honest, or whether a registry holds a receipt as finalized.

The settlement link does read live chain state, and these are its real edges:

• It searches a window around the receipt's issuedAt, not the whole chain, because the public Base Sepolia RPC caps eth_getLogs at 200 blocks per call (measured, not read in documentation). The window it searched is always printed, so "not found" reads as "not in this window" and never as "does not exist". Widen it with --window=.

  A receipt may be signed before its settlement lands, since the seller holds the nonce first, so the window is centred rather than backward-looking. What that costs, measured against a live receipt, with the reach derived from the measured block time:

  --window	reach around issuedAt	RPC calls	wall clock
  50	1.7 min	1	0.7s
  200	6.7 min	3	0.8s
  400 (default)	13 min	5	1.3s
  1000	33 min	11	1.7s
  2000	67 min	21	2.9s

  Calls grow linearly with the window and not with the payer's history, which is the trade the window buys: an hour of tolerance for 21 requests. Searching a payer's full set of consumed nonces instead needs no issuedAt at all and cannot miss on skew, at a cost that grows with how much that payer has ever spent.

• A refused RPC call raises an error instead of returning an empty result. A search that could not run must never be reported as a search that found nothing.

• Only eip3009 is implemented, because that is the only method the specification currently defines.

• It confirms that a receipt commits to one settlement. It does not confirm that the amount or the payee were the agreed ones; read the matching transaction for that.

• Receipt version 2 is defined in a pull request that is still open, so its preimage can still change. The commit it was built against is pinned in the source and in the status table above.

Provenance and license

Apache-2.0, the same licence both verified projects use. Their published test material is redistributed unmodified with attribution; see NOTICE for the exact commits. Neither project has reviewed or endorsed this repository.

Check your own endpoint, in one command

No install, no account, nothing published to npm:

npx github:holistis/agent-receipt-verify https://your-endpoint.example/paid

The ladder itself is a versioned file, profiles/x402-deployed-ladder-v1.json, which names the clause behind every level, the input each check needs, and what to report when a check cannot decide. The tool reads that file rather than carrying the rules in code, so a second implementation can measure the same thing without reading ours.

It measures levels 0 to 2 of the ladder in CONFORMANCE.md without paying. Levels 3 to 5 need a real settlement and come back as not measured, never as failed. What it deliberately does not measure is listed in that same file.

A badge for an endpoint that was measured

If an edition covered your endpoint, there is a badge for it:

[![x402 receipt](https://api.tokenizen.nl/api/x402/badge?url=https%3A%2F%2Fyour-endpoint.example%2Fpaid)](https://github.com/holistis/agent-receipt-verify/blob/main/CONFORMANCE.md)

The same data as JSON is at /api/x402/badge.json?url=....

Two things about that badge, because both are deliberate. It never measures live: it answers from a dated measurement, so nobody can use that endpoint to probe arbitrary hosts, and the number does not change on every hiccup. And the measurement date is rendered inside the badge, so a badge that has gone stale is visibly old rather than quietly wrong.

An endpoint no edition has covered comes back as measured: false, not as level 0. Those are different claims: one says we looked and found nothing, the other says we never looked.