Fairway florist concierge

Fairway Florist is a verified Flary integration. Its Astro storefront embeds Daisy, a durable flower concierge. A Hono route proxies the public chat to a separate Flary Worker through a private Cloudflare service binding.

The storefront owns Shopify, checkout, delivery rules, and customer UI. Flary owns durable threads, model execution, typed tools, requested input, realtime events, and reconnect.

Product boundary

Keep commerce rules in trusted tools

Wrap the exact operations that the agent needs. Keep Shopify access tokens and delivery-provider credentials in Worker secrets or a Flary connection. Never put them in the prompt or tool input.

import { flary, z } from "flary";

const app = flary({
  name: "fairway-daisy",
  applicationId: "fairway-storefront",
  model: "google/gemini-3.7-flash",
  auth: ({ request, bindings }) =>
    verifyPrivateStorefrontRequest(request, bindings.FLARY_INTERNAL_TOKEN),
});

const searchCatalog = app.fn({
  description: "Find available flower products by occasion, budget, or color.",
  input: z.object({
    query: z.string().min(1),
    maximumPrice: z.number().positive().optional(),
  }),
  output: z.array(ProductSummary),
  policy: {
    operation: "read",
    capabilities: ["shopify.products.read"],
  },
  run: (input) => catalog.search(input),
});

const checkDelivery = app.fn({
  description: "Check delivery availability and price for a ZIP code and date.",
  input: z.object({
    zipCode: z.string().regex(/^\d{5}$/),
  }),
  output: DeliveryQuote,
  policy: {
    operation: "read",
    capabilities: ["delivery.quote"],
  },
  run: (input) => fulfillment.checkArea(input),
});

The current customer agent has two read tools: search_catalog and check_delivery_area. Shopify Admin writes are not in the customer agent. This keeps checkout and product publication outside model control.

Add a durable store agent

const daisyTools = app.tools({
  catalog: searchCatalog,
  delivery: checkDelivery,
});

export const daisy = app.agent({
  name: "daisy",
  model: "google/gemini-3.7-flash",
  instructions: `
    Help customers choose flowers and understand delivery.
    Use the catalog before naming or pricing a product.
    Use delivery before stating how a ZIP code is served.
    Use request_user_input when one important detail is missing.
    Never invent availability, prices, delivery promises, or order state.
  `,
  tools: daisyTools,
  eagerTools: ["catalog", "delivery"],
  delegation: { mode: "disabled" },
  limits: { steps: 12, toolCalls: 16, timeoutMs: 90_000 },
});

export default app.serve({ daisy });

request_user_input lets Daisy ask for an occasion, budget, ZIP code, or delivery date and continue the same turn after the answer. Provider and Shopify credentials stay in Worker secrets. They do not enter the prompt.

The browser stores only the durable thread ID. An HttpOnly identity cookie scopes that thread through the trusted storefront proxy. The React client uses flary/client, useFlaryThread, FlaryMarkdown, and FlaryUserInput.

The Flary Worker uses Durable Object SQLite for the live transcript and D1/R2 bindings for the catalog and session archive. The storefront can disconnect, reload, and reopen the same Daisy conversation.

This is the correct pattern for a public storefront: keep commerce authority in the application, expose small trusted reads to the agent, and use Flary for the durable conversation runtime.