Abstract
Agent Rail is a machine-readable interface between software intent and physical production. It translates a structured request — product, quantity, material, finish, artwork, turnaround, destination — into a production-safe order model, then returns deterministic quotes, payment instructions, and normalized status events.
The current deployment is a controlled pilot. Catalog, quoting, order intake, payment-request generation, and status persistence are functional. Production pricing, spot-rate, and end-to-end production submission are deployment-gated: they run against real services when credentials are configured, otherwise they return clearly labeled demo values.
Scope note
01. Complex production lacks a common machine interface
Standardized products are already available through print and fulfillment APIs. The harder problem is complex commercial production, where specifications, materials, artwork, pricing, approvals, manufacturing constraints, shipping, and exception handling vary by product and supplier. Agent Rail is designed to translate structured software requests into production-safe workflows.
02. Agent protocols reduce integration friction
Model Context Protocol (MCP) and emerging agent-communication patterns give an external agent a standard way to discover tools, submit arguments, and receive typed results. Agent Rail exposes catalog, quote, and order tools through MCP-over-HTTP and through a parallel REST surface. These protocols reduce integration friction; they do not create demand or guarantee category leadership.
03. System overview
The web tier exposes public routes and authenticated API endpoints. The quote layer computes pricing and converts CAD to USDC base units. The order layer mints order ids, persists fulfillment payloads, and returns payment requests. The payment layer generates settlement instructions and listens for confirmation. The fulfillment layer maps product lines to production API calls and normalizes status events. The treasury layer sweeps and converts independently.
| Feature | Status | Notes |
|---|---|---|
| Catalog endpoint | PILOT | Public catalog returns configured product lines and the POD price list. |
| Quote endpoint | PILOT | Real-time quote validates the spec and returns a priced response; production pricing config is deployment-gated. |
| CAD pricing source | DEMO | Production CAD pricing client is used only when ADVERTEK_PRICING_API_URL is set; otherwise falls back to inspection mocks. |
| CAD-to-USDC spot-rate source | DEMO | Live spot-rate client is used only when SPOT_RATE_API_URL is set; otherwise returns a fixed inspection value. |
| MCP endpoint | PILOT | MCP-over-HTTP endpoint is available behind API-key auth; underlying pricing follows deployment config. |
| create_order tool | PILOT | Order intake validates, prices, persists, and mints a payable request; production submission is gated by config. |
| Payment request generation | PILOT | USDC settlement instructions are computed server-side; the wallet controls the transfer. |
| Solana transfer confirmation | DEMO | Confirmation polling and memo parsing are implemented; mainnet end-to-end verification is pending. |
| QuickNode webhook verification | PLANNED | Webhook handler is stubbed; signature verification and production stream are not yet verified. |
| Order persistence | PILOT | Orders and status events are stored in the configured Postgres backend. |
| Advertek production API submission | PILOT | Fulfillment order builder maps product lines and POSTs to the production API when credentials are configured. |
| Production status webhooks | PILOT | Advertek status mapping and outbound webhooks are implemented; production event stream is pilot-gated. |
| Treasury sweep | PLANNED | Sweep worker is available for deployment but not wired into the public web tier. |
| OKX conversion and reconciliation | PLANNED | OKX deposit, conversion, and fiat allocation logic is implemented in the treasury worker; not production-verified. |
04. Order lifecycle: a structured path with explicit exception gates
Orders move through validation, quoting, approval, settlement, production submission, and shipment. The system supports four paths:
- Clean: spec validates and the order follows the automated workflow.
- Assisted: artwork, engineering, compliance, capacity, or commercial rules require named human review.
- Rejected: the system returns structured failure reasons and next actions.
- Failed: an unexpected error occurs and is logged for incident review.
05. Settlement architecture
The rail is intentionally payment-rail agnostic. Today it generates a USDC payment request on Solana as a baseline integration: the order id is encoded in the memo, and the payer wallet authorizes the transfer. A US-only fiat on-ramp (MoonPay Buy) lets a card buyer settle on the same rail: the signed widget delivers the invoice's exact USDC to the settlement wallet, and the resulting on-chain transfer is re-verified before the order is marked paid. It is built but not production-verified, and it does not serve Canada. Escrow and marketplace routes remain roadmap work. Treasury sweep, OKX conversion, and fiat reconciliation are implemented in an independent worker and are not production-verified.
06. Security and engineering invariants
- Money is stored and transmitted as integer minor units; floats never cross a trust boundary.
- Zod validates every API, MCP, webhook, and vendor payload.
- Secrets are separated: the web tier holds no money-moving keys.
- Webhook signatures are verified where the protocol supports it.
- Rate limiting and per-IP abuse controls are active on public mutation endpoints.
- Audit logs capture quote, order, payment, and status events without exposing PII.
07. Market context
Agencies and creative operations are the strongest initial wedge. They manage repeat physical output across many clients, each with different brand rules, specs, and approval chains. Standardized print-on-demand and direct-mail APIs already exist; Agent Rail is not differentiated by API access alone. The intended difference is machine-readable access to broader and more complex commercial production.
The defensibility goal is production ontology, pricing logic, job data, exception handling, workflow integration, and network performance — built against a real operating production floor.
08. Roadmap
Verify catalog, specification, quote, and demo flows.
Complete a controlled end-to-end pilot with human exception review.
Add repeat ordering, webhooks, enterprise controls, and measured service levels.
Add qualified production endpoints after repeat demand and routing data exist.
09. Measurable product indicators
- Quote response time
- Specification validation rate
- Human intervention rate
- Quote-to-order conversion
- Production error rate
- On-time shipment rate
- Reprint rate
- Repeat-order rate
- Integration activation rate
10. Changelog
- 2026-09-10
- Added a US-only MoonPay Buy on-ramp that settles onto the existing USDC rail; marked it built but not production-verified.
- 2026-09-02
- v0.2.0 — Repositioned site as agentic production infrastructure. Separated current product status from target architecture.
- Earlier
- Initial technical whitepaper and API surface.
For questions or pilot access, use the form below.
Request Access