---
title: "Nook vs building your own agent shopping integration"
description: "Thinking about letting an agent shop by driving store websites yourself? Compare that path with Nook: real checkout quotes, encrypted sessions, required approval, and one API across stores."
canonical: https://nookstore.xyz/blog/nook-vs-diy-agent-shopping
published: 2026-11-05
last_updated: 2026-11-05
image: https://nookstore.xyz/blog/nook-vs-diy-agent-shopping-og.jpg
image_alt: "Tangled cables and loose browser windows on cracked asphalt beside a tidy Approved Checkout card with Nook's orange starburst and blue eyes."
related:
  - how-ai-agents-checkout-online
  - best-stores-for-ai-agent-shopping
  - what-is-checkout-bench
  - ai-agent-shopping-faq
---

# Nook vs building your own agent shopping integration

**TL;DR.** You can teach an agent to click through a store yourself. You will spend that work on sign-in, cart cleanup, live totals, approval, and the failure cases that make a purchase unsafe to retry. Nook already does that against the store's own checkout. The person connects their accounts once. The agent uses a partner key and a connection link, shows a real quote, and places the order only after the person approves. Build your own interface and product logic; use Nook for the checkout you would otherwise keep rebuilding.

## The decision

If you are adding shopping to an agent, the tempting design is to open the store in a browser, reuse a saved login, and click Buy. It looks simpler than an API because the store already has a website. The costly part is not the first successful click-through. It is getting a total the store will actually charge, keeping passwords and sessions out of the agent, leaving the existing cart alone, requiring a real yes to a quoted price, and handling a changed total, a robot check, or an order you cannot confirm.

Nook is built for that last mile. This page compares the two paths without pretending custom checkout can never work.

## At a glance

| Decision | Building it yourself | Nook |
| --- | --- | --- |
| Where the price comes from | Often a search result, product page, or a parsed cart | The store's own checkout: product, quantity, each charge, and total |
| Credentials | Easy to end up with cookies, passwords, or a shared browser profile in the agent | The person signs in through Nook; the agent never sees what they type |
| Approval | Easy to treat "order this" as permission to buy | The opening request is never the yes; the person must approve the quoted total |
| Existing cart | Easy to check out leftover items, or leave new ones behind | Nook reads checkout and puts the cart back as it was |
| Price changes | You have to notice, and decide whether to buy anyway | Nook buys at the quoted total or less; a higher total, or a changed item, address, or card, stops the purchase |
| Uncertain orders | A timeout invites a second click, and a second order | Nook does not place it a second time; the agent tells the person to check the store's order history |
| Stores | One automation per site, each with its own quirks | Amazon, Uber Eats, Instacart, Walmart, Target, and Chewy can quote and buy today; more stores can connect |
| Interface | Custom browser code | One partner-key API, REST or MCP |
| Cost to the buyer | Whatever you pass through | Nook adds nothing to a checkout; the person pays the store the total they approved |

## Real checkout figures versus a scraped price

A listing price answers "what does this product show as?" A checkout total answers "what will this account be charged, at this address, with this card, right now?"

Those are different numbers. Shipping, tax, tips, bag fees, promotions, and marketplace sellers show up at checkout. If your agent reports a search-result price as the price, it is not quoting checkout.

Nook reads the store's checkout page and returns the charge rows, a total those rows add up to, and the destination and payment it matched. The agent should repeat those figures and nothing extra. The API, MCP tools, and preferences exist to get that quote in front of the person and to buy it only after they agree.

See [How AI agents check out online](/blog/how-ai-agents-checkout-online) for why checkout is the hard part.

## Credentials and sessions

A homemade integration has to solve sign-in, and the risky shortcuts are asking for a password in chat, reusing a logged-in browser profile, or storing cookies next to the agent's tools. Nook keeps that work on the accounts side:

- The person signs in to Nook with their phone.
- They sign in to each store in Nook's own form.
- Saved store sessions are encrypted.
- A connection counts only after the saved sign-in works in a brand-new cloud browser.
- The agent receives a `connection_id` it can quote against, not the password and not the session.

A partner key identifies *your* agent to Nook. It is not a buyer identity. Keep it in your host's private storage. Never show it to the person or ask them for one. You can quote only on a connection the person made through your own link.

The person can see and disconnect every connection at [nookstore.xyz/v3/accounts](https://nookstore.xyz/v3/accounts).

## Approval is part of the protocol

If you build checkout yourself, you also have to invent a rule for when the agent is allowed to press the final button. Vague rules produce surprise orders.

Nook's rule is narrow and consistent:

- "Order me X" means the person wants to buy, so the agent should get a quote and ask.
- Their yes is confirmation of *that* total, in whatever clear words they use.
- Their opening request is never the yes, even when it says "order", "buy", or "get".
- Without those words, they approve with Buy on Nook's signed-in review page.
- The person can cancel on that page until Nook starts placing the order.

Pass their words through exactly. Do not paraphrase an approval or approve on their behalf.

## Cart hygiene and store-specific behavior

People already have items in Amazon or Chewy carts, and stores have their own quantity limits, pickup flows, and option pickers. A script that "just adds to cart and checks out" can purchase the wrong basket or leave the cart changed. Nook's documented behavior includes work you would otherwise maintain:

- **Amazon:** only the lines Nook added are checked out; the rest of the cart stays. For a general request it picks the cheapest listing that fits, and asks the person when listings differ.
- **Uber Eats:** one restaurant per order; required item options are asked, never guessed; tips can be requested; order status can be refreshed after confirmation.
- **Instacart:** one retailer per checkout quote, or a cross-store search first so the person can pick a listing.
- **Walmart and Target:** shipping or pickup; the selected card is checked, not switched; quantity limits follow the store.
- **Chewy:** always a one-time purchase (Buy once, not Autoship); existing cart items are set aside and brought back.

Your agent still has to explain choices and warnings. It does not have to reverse-engineer each cart. See [stores.md](https://nookstore.xyz/stores.md) and the [agent guide](https://nookstore.xyz/agents.md) for current limits.

## Failures you cannot hand-wave

**The total changed.** Nook reads checkout again before it buys and stops if the total went up, or if the item, address, or card changed.

**The run did not finish cleanly.** Nook returns blockers and warnings instead of silently dropping a requirement. If the store's quotes cannot honor something — a delivery window, for example — the agent should say so.

**The confirmation is unreadable.** If Nook pressed the final button but could not read the result, the purchase is uncertain. The agent must tell the person to check the store's order history and must not buy again.

**The account is busy.** A connection has one browser at a time. A second quote or purchase waits or comes back as in use, rather than opening another hidden checkout.

Building this recovery yourself is possible — and most of the project.

## What you still build

Nook does not replace your agent. You still decide how people talk to it, which products you suggest, how you present listings, and how you keep each buyer's connection IDs bound to that buyer. Follow the request rules: put the product's own words in the product field, put other requirements in the fields that exist, and never invent a budget, address, or card the person did not provide.

MCP clients can add `https://nookstore.xyz/mcp` with the partner key as a Bearer token; the tools take the same bodies as the API.

[Checkout Bench](/blog/what-is-checkout-bench) grades models on these real checkouts. Measured scores are not published yet. It is not a substitute for your own evaluation.

## When building it yourself can still be the right call

Use a custom integration if you only need one internal tool, on one store you control, with no consumer passwords involved, and you will maintain the click path. Nook is the better default when the person already has accounts at Amazon, Uber Eats, Instacart, Walmart, Target, or Chewy; when several agents should shop for the same person; when credentials should stay out of the agent; and when the person must approve a quote before anything is purchased.

Nook is in private beta. There is no public paid plan to drop into a build-versus-buy spreadsheet. Connecting accounts is free for buyers during the beta, and partner registration is free for agents. When pricing exists, it will be published on [Nook's pricing page](https://nookstore.xyz/pricing.md).

## Get started

**If you are building an agent:** follow the [quickstart](https://nookstore.xyz/docs.md), then keep the [agent guide](https://nookstore.xyz/agents.md) nearby for store fields, polling, and recovery. Register a partner key, send a connection link, and buy only after the person approves the total.

**If you want an agent to shop for you:** connect your stores at [nookstore.xyz/v3/accounts](https://nookstore.xyz/v3/accounts). Sign in with your phone.

For store-by-store fit, see [Best stores for AI agent shopping](/blog/best-stores-for-ai-agent-shopping). For safety questions, see the [AI agent shopping FAQ](/blog/ai-agent-shopping-faq).