PreFlight

A service can be online and still be unbuyable

Health checks do not prove that a paid agent service works for a real customer.

A service can respond successfully while its live contract advertises the wrong price, network, asset, or payee. Payment can settle while delivery fails. Declared behavior can drift from production. A generic score can hide the exact issue blocking release.

PreFlight tests the buyer journey before customers discover the failure.

• Contract drift between what was declared and what is live

• Incorrect payment challenge terms

• Successful settlement followed by failed delivery

• Duplicate-payment or replay inconsistencies

• Missing evidence and unclear remediation

Related MCP server: Tilde x402 MCP Server

How PreFlight works

PreFlight separates free discovery from paid release verification. Nothing is charged until the developer confirms the discovered contract and starts the full check.

1. Submit a public endpoint or an OKX.AI Agent ID.

2. Discover the live service contract.

3. Review a proposed manifest with per-field provenance.

4. Confirm the release check and complete the x402 payment.

5. Optionally authorize bounded buyer proof against the target service.

6. Receive a decision, criterion evidence, remediation, and a signed receipt.

What PreFlight verifies

Decisions you can ship against

Each decision is built from criterion-level states:

• MATCH -- observed production behavior matches the declaration

• CONTRADICTION -- observed production behavior conflicts with the declaration

• UNKNOWN -- there is not enough safe evidence

• NOT_APPLICABLE -- the criterion does not apply to this surface

Real product output: verdict, criterion evidence, remediation, and a PreFlight Signed Receipt.

Key capabilities

• Discovery-first input -- start with a public endpoint or an OKX.AI Agent ID rather than a hand-authored manifest.

• Agent ID resolution -- pass a numeric Agent ID and PreFlight resolves the listing to its live endpoint, pre-filling manifest fields with observed provenance.

• Proposed manifest with provenance -- inferred fields show where they came from and whether confirmation is required.

• Bounded buyer proof -- with owner attestation and spend limits, PreFlight can pay and take delivery like a real customer.

• Deterministic verdicts -- RELEASE, BLOCK, or UNKNOWN from criterion states, not an opaque aggregate score.

• Actionable remediation -- contradictions include observed evidence and the exact issue to fix.

• Signed receipts -- Ed25519 signatures over canonical JSON, verifiable outside the report page.

• Public receipt verifier -- anyone can verify a receipt at /verify without an account or token.

• Cohort scan -- free discovery runs across every listed OKX.AI ASP, with conforming services named and contradictions reported as codes and counts only.

• Per-ASP permalinks -- each agent has a permalink at /asp/{agent_id} showing its runtime evidence state.

• Passport -- owner-authorized, scoped release passport tied to a receipt and policy version.

• Badge embed -- 88x28 SVG badge at /api/v1/badge/{agent_id}.svg for services with an active passport.

• Benchmark corpus -- adversarial fixtures with seeded faults, tested against the current policy; failing fixtures render as failing.

• Self-check -- operator-funded dogfooding verification with customer_demand: false, published for transparency.

• Private reports -- bearer capability tokens protect non-public report data.

• Agent-native access -- use the hosted API, MCP endpoint, or npm CLI.

Use PreFlight your way

Web

Open usepreflight.xyz, paste a public endpoint or an OKX.AI Agent ID, and review the proposed manifest. Discovery is free. A full release verification is 0.10 USDT over x402.

API

curl -s https://api.usepreflight.xyz/api/v1/service | jq
curl -s https://api.usepreflight.xyz/api/v1/contracts/verify-release-request/v1 | jq
curl -s https://api.usepreflight.xyz/api/v1/contracts/release-manifest/v1 | jq

POST /api/v1/verify-release is the paid x402 release-check endpoint. An unpaid request returns a payment challenge; a funded client replays the request with the required payment proof.

Canonical minimum paid request:

{ "endpoint": "https://public-service.example/path" }

schema_version is optional and defaults internally. agent_id is the supported alternative to endpoint; provide exactly one target. Idempotency-Key is optional for generic buyers.

Public endpoints (no auth required):

See docs/api.md.

MCP

{
  "mcpServers": {
    "preflight": {
      "url": "https://api.usepreflight.xyz/mcp"
    }
  }
}

CLI installation is not required for the hosted MCP service.

See docs/mcp.md.

CLI

npm install -g @vinaystwt/preflight-cli
preflight --help
preflight verify --help
preflight verify-receipt --help
npx @vinaystwt/preflight-cli --help

See docs/cli.md.

Architecture

PreFlight keeps discovery, payment verification, bounded buyer execution, criterion evaluation, report storage, and receipt signing as explicit trust boundaries.

Web, CLI, and MCP clients call the Fastify API. Discovery uses guarded egress to inspect public services. The seller gate verifies payment to PreFlight, while bounded buyer execution can independently pay the target service. A deterministic criterion engine writes private reports and signs portable receipts with Ed25519.

See docs/architecture.md.

PreFlight Signed Receipts

Every completed verification issues a portable receipt. Anyone can verify that this receipt was issued by PreFlight, has not been altered, and applies to the identified runtime snapshot and policy version. It does not establish that PreFlight observed correctly, that the policy is correct, or that a target is safe.

• Canonical JSON with sorted keys

• SHA-256 payload hash

• Ed25519 signature

• Public verification keys at GET /api/v1/pubkeys

• Public receipt lookup at GET /api/v1/receipts/{receipt_id}

• Public verifier at usepreflight.xyz/verify

See docs/receipts.md and docs/cli.md.

Security and trust boundaries

PreFlight is a verifier, not a custody product.

• Private reports require bearer capability tokens.

• Reports are published only after the required settlement state.

• Probe egress rejects private/internal targets and unsafe redirects.

• Bounded buyer proof requires owner attestation and spend caps.

• Terms-hash drift aborts buyer proof before payment.

• Seller, buyer, and operational wallets are separated by role.

• Ambiguous evidence returns UNKNOWN rather than a false release.

See docs/security.md.

Backend:

npm ci
npm run typecheck
npm run build
npm test -- --run
npm run migrate
npm start

Web:

cd web
npm ci
npm run lint
npm run build
npx tsc --noEmit
npx playwright test --reporter=list

CLI:

npm run build --prefix packages/cli
(cd packages/cli && npm pack --dry-run)

src/                 Fastify API, MCP, release criteria, payments, receipts
src/db/migrations/   Additive database migrations
web/                 Next.js web application
packages/cli/        Published CLI source
docs/                API, MCP, receipt, security, architecture, and deployment docs
test/                Backend and release-gate tests

Status and limitations

• Agent ID resolution is intentionally conservative unless an authoritative listing resolver is configured.

• Gallery entries are opt-in.

• Chain anchoring is disabled unless explicitly configured and proven.

• Browser receipt verification may fall back honestly if local Ed25519 support is unavailable.

License

No public license has been selected. Do not assume rights to reuse, modify, or redistribute the source unless a license is added.