ReceiptRail

ReceiptRail (formerly AgentToll) — on-chain delivery receipts & official x402 receipt verification on Solana

⚠️ Official project notice / 身份声明: This is the official ReceiptRail repository, maintained by GitHub org xka0085-byte and npm account eidonze. We are NOT affiliated with agenttoll.dev, npm/agent-toll, agenttoll-mcp (Base), @agenttoll/sdk, 402.ad/AgentToll, or the "Receiptrail" accounting SaaS. Verify you are talking to this project by checking: MCP endpoint https://agenttoll-receipts.app.workbuddy.host/mcp · Official MCP Registry id io.github.xka0085-byte/receiptrail · npm package receiptrail-mcp (published by eidonze).

🔎 Early access (2 slots): I'll manually inspect your x402 endpoint's

payment-to-delivery path and send you a conformance report within 24h — $9/$19.

Details: https://x402-endpoint-inspection.app.workbuddy.host/

ReceiptRail is an on-chain delivery-receipt service for x402 AI-agent payments on Solana: after a settlement, sellers anchor a SHA-256 digest of the delivered content plus the settlement reference into a public PDA. Anyone — the buyer agent, an auditor, a regulator — can independently verify what was delivered, without trusting the vendor, the buyer, or any API we operate.

⚡ Live service (use it now)

🆕 verify_x402_receipt — official x402 receipt-format verification (v0.3)

The x402 Foundation now ships a standard Signed Offers & Receipts extension (docs.x402.org/extensions/offer-receipt, npm @x402/extensions): servers sign an offer on every 402 and a receipt on every 200 — portable proof-of-interaction artifacts for reputation, auditing and disputes.

ReceiptRail is a Solana-side verification node for that format: our MCP tool verify_x402_receipt validates any official-format receipt artifact (JWS with EdDSA/Ed25519 — Solana-native — or ES256/P-256), resolves the signer key automatically from did:key / did:web / did:jwk, enforces freshness and payload assertions, and anchors verification results on-chain. Cross-verified against the official @x402/extensions verifier (same verdicts on valid and tampered artifacts — see receipts-service/test/cross-verify.mjs).

Keywords for agent discovery: x402 receipt, x402 offer-receipt, signed receipt verification, x402 dispute, delivery receipt, proof of delivery, proof of payment, Solana, USDC, MCP, AI-agent payments, verifiable settlement, agent reputation, on-chain audit.

Agent / Chain Evidence Tools — the suite

ReceiptRail is the flagship of a six-tool evidence suite (all read-only, no keys, JSON output):

Live tools page: https://x402-endpoint-inspection.app.workbuddy.host/tools.html

---

AgentToll is the delivery-receipt protocol behind ReceiptRail. A vendor pins the SHA-256 digest of a report's exact bytes into a public PDA account on-chain. Anyone can independently recompute the digest from the delivered file and compare it against the chain, without trusting anyone.

"Don't trust what we say. Run the command."

Every capability claim in this README maps to a command in this repo and, where possible, a Solscan link on devnet.

Status: W1–W3 complete, independently audited (reports in audit/). Deployed on Solana devnet.

• Program ID: 8ACN1KNEFXM2N2FMzxTfAzB1c3g5n47ZuhbPoUXPnCTp — view on Solscan
• Receipt created by the recorded demo payment flow: delivery tx · payment tx

Honesty note: the demo uses test-USDC, a self-issued devnet mint with no real-world value. It is not real USDC. We say so everywhere, including in code output.

Why this exists

Machine-to-machine payments (the x402 pattern) are becoming real: agents pay per API call. But after an agent pays, it holds no trustworthy proof of *what* it received. Receipts, invoices, or hashes served by the vendor itself are indistinguishable from fabricated ones. AgentToll moves that proof to a neutral public ledger: the digest is written once, is tamper-evident (same id + different digest is rejected on-chain), and is verifiable offline forever.

Verify an existing receipt right now (no vendor involved)

cd verifier && npm install && cd ..

node verifier/verify.mjs \ --url https://api.devnet.solana.com \ --vendor B7wGaKwEGAmNP7PEhqwFrvfCQPH4zwiE7FZNLoeB4naD \ --report-id x402-demo-001 \ --file demo/report-402.json

Expected: JSON with "verdict": "PASS" and all five checks true (C1 ownership, C2 PDA derivation, C3 schema, C4 content binding, C5 freshness). Exit code 0. Then tamper with the file (add one byte) and rerun — C4 fails, exit code 1. That mismatch is the product.

Architecture

sequenceDiagram

participant B as Buyer agent participant V as Vendor HTTP service participant S as Solana devnet participant C as Offline verifier B->>V: GET /report/x402-demo-001 V-->>B: HTTP 402 + payment challenge B->>S: SPL transferChecked (test-USDC) B->>V: POST + X-Payment-Signature V->>S: getParsedTransaction, verify payment V->>S: create_receipt(report_id, SHA-256) V-->>B: report bytes + receipt metadata B->>C: Spawn verifier with local report C->>S: Read receipt PDA via public RPC C-->>B: C1-C5 PASS/FAIL JSON

The on-chain program enforces the anti-tamper rule at the ledger level: a PDA seeded by (vendor, report_id) can be created once; re-creating with the same digest is idempotent, with a different digest fails with DigestConflict. Overwrites are impossible, not just discouraged.

Run the full x402 demo

Requires Node.js 22+, a devnet-funded vendor keypair, and the deployed program above.

# 0. install the single JS dependency set

cd verifier && npm install && cd ..

1. configure (bash example)

export AGENTTOLL_RPC_URL='https://api.devnet.solana.com' export AGENTTOLL_KEYPAIR='/absolute/path/to/vendor-keypair.json'

2. create test-USDC mint, fund buyer (idempotent)

node demo/setup.mjs

3. start vendor (localhost:8787)

node demo/vendor.mjs

4. unpaid request must be refused

curl -i http://127.0.0.1:8787/report/x402-demo-001 # → HTTP 402

5. run the buyer agent: pay → receive report+receipt → verify 5/5

node demo/buyer-agent.mjs

6. forged payment signatures must be rejected

curl -i -X POST -H 'X-Payment-Signature: not-a-real-signature' \ http://127.0.0.1:8787/report/x402-demo-001 # → HTTP 402, no report

demo/buyer-agent.mjs prints {payment_sig, receipt_tx, pda, verifier_verdict, checks} and exits 0 only if the spawned offline verifier passes all five checks.

On-chain program (W1)

Anchor 1.2.0, two instructions:

• create_receipt(report_id, digest) — first call creates the PDA; same id + same digest is idempotent; same id + different digest → DigestConflict
• verify_receipt(report_id, digest) — permissionless; emits a ReceiptVerified event

Build and test locally:

./build.sh        # = anchor build --arch v0

cargo test --all # 8 litesvm tests, incl. DigestConflict & idempotency

Audits

Every milestone was reviewed by an independent auditor that did not write the business code. Reports with command outputs and on-chain evidence:

• audit/W1-2026-09-14.md — contract, 8/8 tests, devnet deployment
• audit/W2-2026-09-14.md — offline verifier, tamper-detection e2e
• audit/W3-2026-09-14.md — x402 flow, negative paths, idempotent replay

Scope & limitations (read before trusting anything)

• test-USDC is self-issued on devnet; it stands in for a real stablecoin. No real value.
• The demo serves one report id on localhost; it is a protocol demo, not a hosted service.
• The on-chain receipt binds report_id + digest; the payment signature is linked via the demo flow, not stored on-chain (the frozen W1 schema has no such field). See W3 audit §findings.
• A professional third-party security audit is on the roadmap and has not happened yet.

Disclosed prior work

This project was built for the Colosseum CWF hackathon. Per the rules, we disclose pre-existing development: t3n-recon-agent and z-tenant-recon — a TEE-based prototype implementing the same anti-tamper digest logic (same-id-different-digest rejection, independent verifier), from which the on-chain design was ported. All AgentToll commits, tests, and deployments happened during the competition window (first commit: 2026-09-14).

中文说明见 README.zh-CN.md。

---

Suite hub

Part of the Agent / Chain Evidence Tools suite — read-only, no-keys, no-payments diagnostics for AI agents on Web3.