Case Study: How an AI Agent Buys a Product in 8 Steps (Demo Store Walkthrough)

What does it look like when an AI agent buys something? In this walkthrough, we'll follow a Claude agent through every step of a sandbox purchase from the Trusteed demo store: 18 products, 6 categories, 3 discount codes, multiple shipping options, and a checkout flow that handles payment and address collection natively inside the chat. Every curl command in this post works against the live demo endpoint right now — and the endpoint itself reports `store_environment: "sandbox"` on every search response. The final step uses `payment_method: "MOCK"`: it creates an order and moves no money. Which rails can settle real value depends on the rail and on each merchant; /en/status carries the per-rail state taken from the signed capability registry. Correction 2026-10-11: an earlier version of this paragraph denied that any payment settles on the platform; that was a dated statement nobody re-measured, and the state is per rail and per merchant.

The demo store ships with 18 curated products across 6 categories: sneakers (4), jackets (3), accessories (3), pants (3), t-shirts (3), and footwear (2). These aren't placeholder items: each product has realistic pricing, variant options (size/color), stock levels, ratings with review counts, and complete metadata including weight, material, and origin. Ratings are merchant-asserted, and every search response says so in a `rating_provenance` block — Trusteed does not verify them. Key products you'll see in this walkthrough (measured against the live endpoint on 2026-09-22): • Nike Air Max 90: $130 (was $160), 4.7★ from 324 reviews, sizes US 8–10 • The North Face Nuptse Jacket: $320 (was $380), 4.8★ from 512 reviews, 700-fill goose down • Adidas Ultraboost 23: $190, 4.5★ from 189 reviews • New Balance 574 Classic: $89.99, 4.3★ from 156 reviews The demo store's trust score was 44.7/100 (v4.1), band RESTRICTED, STANDARD verification when this post was last reviewed on 2026-09-22. Don't take that number from here: it moves, and the authoritative value is signed and served at https://api.trusteed.xyz/api/v1/stores/demo-store/trust-score. Every MCP response from this store carries the current value in its `merchant_trust` block. RESTRICTED clears the hard eligibility floor (≥ 0.3 on the 0–1 scale) — so the agent can transact — but it is not a high score, and it is not meant to be: the demo store is a sandbox with no settled orders behind it, and the model says so.

The agent starts by searching the catalog. No authentication needed. ```bash curl https://trusteed.xyz/demo-store/mcp \ -X POST \ -H "Content-Type: application/json" \ -H "Accept: application/json" \ -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"search_products","arguments":{"query":"running shoes","limit":5}},"id":1}' ``` The response returns matching products ranked by relevance and trust score. Each result includes: id, title, price, currency, rating, stock status, and image URL. The agent now has product IDs it can use in subsequent steps.

With a product ID from search, the agent fetches complete details including variants, specifications, and Schema.org JSON-LD. ```bash curl https://trusteed.xyz/demo-store/mcp \ -X POST \ -H "Content-Type: application/json" \ -H "Accept: application/json" \ -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"get_product_details","arguments":{"product_id":"prod-001"}},"id":2}' ``` The response includes everything the agent needs to make a recommendation: price vs compare-at price (for discount calculation), variant options with per-variant stock, material composition, weight, and a full JSON-LD block the agent can use for semantic reasoning.

Now the agent adds items to a cart. This requires a sandbox API key. ```bash # First, get a 24-hour sandbox key curl -X POST https://trusteed.xyz/api/v1/sandbox/key # Returns: {"key":"agnt_sandbox_...","expires_in":86400} # Create cart with the key curl https://trusteed.xyz/demo-store/mcp \ -X POST \ -H "Content-Type: application/json" \ -H "X-Agent-Api-Key: agnt_sandbox_..." \ -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"create_cart","arguments":{"items":[{"product_id":"prod-001","quantity":1,"variant_id":"v-snk-001a"}]}},"id":3}' ``` The response includes a `cart_id` (UUID), line items with pricing, and a `status: "NOT_READY_FOR_PAYMENT"`. The cart needs shipping and checkout to be complete. The `agent_session_id` is also returned, which can be used to resume this cart later.

This is where things get interesting. When the agent calls `get_shipping_rates` without a shipping address, the MCP server triggers elicitation: a native MCP feature that collects user input inline. In Claude.ai, the user sees an inline form appear directly in the chat: - Street address - City - Country code (ISO-2: US, ES, DE...) - ZIP code - Province/State (optional) The user fills in the form, Claude passes the data back to the tool, and shipping rates are calculated. No redirects, no external pages: the entire interaction happens inside the conversation. The demo store returns two options: - Standard (5–7 days): $9.99 - Express (2–3 days): $24.99

Throughout this flow, several invisible systems are at work: • Trust scoring: The demo store's score (44.7/100 on 2026-09-22, band RESTRICTED) clears the hard eligibility filter (≥ 0.3 on the 0–1 scale), which is what lets the checkout proceed at all. Trust also carries 30% of the search ranking weight, so a RESTRICTED store is eligible, not favoured. • Protocol detection: Each request is analyzed by the protocol-detector middleware, which picks an adapter from the payment method. Read "supported" carefully here: the demo store has ACP, AP2, x402 and UCP switched on in its own configuration, but the platform state of those rails is what decides whether anything can settle — and today none of them settles. ACP is deployed with no merchant configured in production; AP2 v0.1 over REST is withdrawn and v0.2 over UCP is still under construction; x402 returns `x402_not_configured` for every merchant. The signed capability registry at /.well-known/trusteed-capabilities.json is the arbiter, and /en/status renders it. That is why this walkthrough uses MOCK. • KYAI policy engine: Every checkout intent passes through the merchant's policy rules before an order is created. What they return depends on the rules that merchant has configured and on the agent's own identity and mandate, not on the store's score alone — a blocked checkout comes back as a refusal with a rule code, not as a silent failure. • Rate limiting: The demo endpoint caps at 20 requests/minute per IP. Sandbox keys get 50 checkout requests/hour. Production access is rolled out according to environment and merchant enablement, not public self-serve tiers. • Session cleanup: Cart sessions expire after 30 minutes of inactivity. Each request creates a fresh MCP server instance: no persistent connections.

The fastest way to experience this flow: 1. In Claude.ai: Add the MCP server URL `https://trusteed.xyz/demo-store/mcp` to your Claude Desktop config and ask Claude to 'search for running shoes on the demo store' 2. Via curl: Copy any of the commands above and run them in your terminal. Discovery tools work immediately with no setup. 3. In the playground: Visit trusteed.xyz/en/demo-store and use the interactive MCP playground to run tools visually with a session trace panel. 4. Via NLWeb: Try natural language search at `POST /demo-store/ask` with `{ "query": "cheap waterproof jacket", "mode": "list" }` for semantic vector search.