cartwright
Features

Agent readiness

Every Cartwright store answers AI agents with a typed, rate-limited, honestly-discoverable public surface — and you can measure yours with a public scorecard.

AI agents — ChatGPT, Claude, Perplexity, shopping copilots, IDE agents — increasingly are the customer's first visit. A Cartwright store doesn't just tolerate them; it gives them a real, documented, safe interface. And because claims are cheap, you can point an independent scanner at your store and get a public scorecard.

Ships in engine v0.46.0. Everything below is profile-aware and flag-gated — a capability your store doesn't run is a genuine 404, never a false advertisement.

Measure it — the public scorecard

Is Agentic is an independent scanner that probes a domain the way an agent would — discovery files, machine-readable errors, typed APIs, trust pages, crawler access — and publishes a scorecard anyone can check.

  • See a live example: the scorecard for our demo store, Northbound Coffee Roasters, running the current engine unmodified.
  • Test your own store: run npx is-agentic your-domain.com or scan it at is-agentic.com — no signup, the result is a shareable public page.

We deliberately don't print a score here: scores move with the scanner's rubric and with every release. The live scorecard is the honest reference — always current, always public, never our word for it.

What an agent gets from your store

A narrow, anonymous read surface. Product search and lookup, category and published-page listing — five read-only tools, rate-limited per IP, over both REST and MCP. Drafts, customers, orders, admin and every write stay behind scoped Bearer keys, and an invalid key never falls back to anonymous access.

A typed contract. OpenAPI 3.1 generated from the same tool registry the store actually runs, one typed operation per tool, with per-operation security — plus an SSR developers page documenting auth, scopes, rate limits and versioning.

Machine-readable errors. Agent-facing errors are RFC 9457 Problem Details (application/problem+json) with a code and a resolution, and rate-limited responses carry RateLimit-* and Retry-After headers — an agent that hits a wall is told which wall and what to do about it.

Trust anchors and markdown. Locale-aware about/privacy/contact pages, correct canonical/hreflang/social metadata resolved from the runtime URL, and Accept: text/markdown content negotiation so text-first agents skip the HTML — including real markdown 404s with recovery links.

Honest discovery. llms.txt, an RFC 9727 API catalog, MCP server cards and a digest-verified public Agent Skill describe only interfaces that exist in your store's profile. White-label gating means a fork never claims to be Cartwright.

The honesty rule

The engine never enables a capability just to score points. OAuth, A2A cards, payment protocols and other advertised surfaces appear in discovery only when they are switched on and working in your store. A scanner probing a disabled surface gets a genuine 404 — which costs points and earns trust. We think that trade is the whole game.

On this page