Amazon is connected, and your agent can check out there.
Get started for free with just a text.
Connect your accountAmazon✓ Connected
Uber Eats✓ Connected
Instacart✓ Connected
WalmartOne tap to connectYour agent sends you one link, and you sign in to Nook with your phone and connect your store accounts in one tap while your agent never sees what you type.
Nook works through your signed-in account, finds the product, and reads the store's own checkout figures.
You see the product, quantity, price, address and payment before anything is bought, and saved preferences fill only what a request leaves out.
You approve in your own words in chat or with Buy on Nook's review page, and Nook presses the store's final order button once and reports the order number back.
Any agent can register its own partner key, and yours connects your accounts in one link.
Muse, ChatGPT, Grok and open-weights models can all finish the checkout as long as they can call a tool, because the connection belongs to the person and the model is only a detail.
Claude58s
ChatGPT1m 12s
Grok2m 03s
Nook
Amazon checkoutVerified just nowNook reads the store's own checkout — product, quantity, price, address and payment — not a search-result guess.
Learn moreYou set the dial per store and per purchase, approving in one word in chat or reviewing every figure on the page first, and nothing is bought until you choose.
Learn moreFast, deterministic settings to configure your agent’s preferences.
Learn moreKeep human and agent accounts stored separately, so your agent never touches your personal logins.
Learn moreEasily create authorized connections, so people rarely revisit passwords and never deal with browsers-in-browsers.
Learn moreSigning in to ChewyEvery quote and purchase leaves a trail of the figures, the order number, the live status or exactly what stopped it, so everything your agent does stays traceable by you however much it shops.
Learn moreSave your usual address, card and place names once in Nook, and whichever agent or AI app you shop through, the same preferences fill the same gaps and show on every quote.
Learn moreYour preferences · ChewyAgents can't change theseApplied only where a request leaves a gap and reported on every quote.
Checkout Bench is an evaluation of how well AI agents shop on real stores. Each task is a request written the way shoppers actually type, such as “can you get me the dji mic mini thing from amazon,” and the agent carries it out inside a real, signed-in store account in a fresh cloud browser that is thrown away after the run. Every run stops before payment, and the final cart is graded in code for the right item, the right quantity and a total under the price cap, along with whether the agent asked when the request was genuinely unclear, so no model is asked to judge another. We built it because shopping agents tend to fail quietly on variants, sign-ins and checkout steps that only a real store exposes, and a benchmark that never reaches those steps cannot measure them.
Checkout Bench runs every model through the same shopping requests on real, signed-in store accounts and grades the final cart in code, and its leaderboard ranks models by how often they finish a request correctly across the six stores that return a full cart and total.
| Rank | Model | Pass rate, all 6 stores |
|---|---|---|
| 1 | Model A Illustrative | 62% |
| 2 | Model B Illustrative | 55% |
| 3 | Model C Illustrative | 47% |
| 4 | Model D Illustrative | 33% |
Checkout Bench measures how well AI agents shop on real stores, and it rests on three decisions about what a task is, how a run is scored and why the result is worth trusting.
A task is one request written the way shoppers type, paired with the products that count as correct or a note that the right move is to ask, and the agent works on it inside a real, signed-in store account in a fresh cloud browser.
Code checks the final cart for the right item, the right quantity and a total under the price cap, then confirms that the run stopped before payment and nothing was bought, so no model is ever asked to grade another model.
Shopping agents rarely fail loudly, because a wrong size, a stalled sign-in or a checkout step they can’t read simply ends the sale, so measuring on real checkouts shows where each model gets stuck before shoppers find out.
The lab is where every Checkout Bench run takes place, a signed-in workspace that sends real shopping requests to AI assistants through Nook, records everything each one did and scores the result against known answers.
Quote suite sends one shopping request to several AI assistants at once, and each one uses Nook to get a real checkout quote from a real, connected store account. You pick the store, the assistants and what the shopper asks, and every run stops before payment.
Every assistant in a run gets its own verdict card with the quote total or the blocker, whether the request matched, and its time, tokens and cost. Its details open onto the requests it sent, the quote or blocker, the conversation, the checks, the findings in order and the full trace.
Overview shows what the lab is doing and what it can reach, with the lab’s status, the assistants on hand, the stores connected through Nook and a fresh sandbox VM for each model, above the most recent runs.
Runs lists every run the lab kept, newest first and grouped by day, with counts of completed and failed runs and the share of clean verdicts, and each row opens, replays or downloads its evidence.
Trace replay opens a single run as “Inside this shopping run,” with what the assistant actually said, the requests it recorded and their outcomes, and a step-by-step timeline you can play, pause and filter by source.
Model comparison sends one request to separate assistants, each in its own VM with real tools and responses, and lays their results side by side so the models can be read against each other.
Bench runs fixed shopping tasks with known right answers, starting from an answer sheet, scoring every attempt as a correct or wrong checkout, and collecting the results in a scorecard with the causes behind each failure.
Accounts lists the stores connected to the lab through Nook and when each one was last verified, next to the stores still available to connect in Nook.
Every request can run as one consistent shopper, with a short written profile that a simulated shopper answers the assistant’s questions from or that the assistant keeps as its own notes. The profile travels with the real store accounts connected through Nook, so the sizes, addresses and choices an assistant asks about come from the same person each time.
Nook keeps the figures from every quote and order, so a month of agent shopping reads like a statement, and the numbers in these charts are illustrative sample data rather than anyone’s real activity.
Paid from banks linked through PlaidOctober, one phone numberCheckout Bench runs real models through Nook against signed-in stores and a replica of Amazon built from 211 real runs, and every figure in this section is its actual result rather than sample data.
Nine of the ten tasks land within 15 points of live Amazon, and the one that misses is the DJI Mic Mini kit, where live Amazon showed three competing listings and the replica's September data had gone stale.
Six of the seven task kinds already score 88 to 100 percent for every model, so lookalike products are where the models pull apart, and GPT-5.6 Sol and Luna avoid wrong quotes by asking the person to choose instead.
The live Amazon column is interim because those batches are still running, so its figures may still shift a few points.
A run can go all the way through an approved payment or end at a verified quote, and these 31 graded runs used quote-only mode, so each tile below is how a run ended.
Checkout Bench turns your site’s own sign-ins, carts and checkouts into graded tasks and fine-tunes a model on the attempts that pass, and reinforcement learning can then reward the runs that ask when a request is unclear, so the model learns how your checkout actually behaves instead of guessing from every other store.
Real requests replay against a replica of your checkout, many tries each.
The same grader checks the product, the address and the budget on every try.
Passing attempts become training data, and the next round starts from the new checkpoint.
Held-out tasks run before and after on your real site, so only the training changes.
These numbers are illustrative and show the shape of a training run rather than a measured result, while the pipeline that produces them already runs on the Amazon replica.
The same cart is rarely the same price everywhere, so Nook quotes it across every store you connected at once, watches the prices you care about and speaks up when one moves.
$86.20$77.80$84.55
Your agent handles the standing errands, like the restock before the bag runs out and the Friday list before the weekend, and it learns the rhythm so each one comes to you as a single approval.
Each store you connect becomes available to whichever agent you authorize, and the connection keeps its owner, status, start date and last order on record, so nothing about your accounts is a black box.
Your phone number is your sign-in, because Nook texts you a code and, once you enter it, your agents can reach every account you connected without passwords to hand out or store-by-store logins.
Approval lives in the conversation you already have with your agent rather than in a separate app, so any clear affirmative reply to the quote’s total, such as “approve”, “confirm”, “buy it” or “I’m down”, counts in your own words. Nook passes your words along exactly as you said them and keeps them with the order, and if you’d rather look first, Buy on Nook’s review page works too because the review doesn’t expire.
nookThe Amazon quote comes to $50.00 for three items, and the reply Approve places the order.
Penne Vodka Pasta from Uber Eats is $23.77 with tip, and buy it is all it takes.
Two seats for Dune: Part Three at AMC tonight cost $31.50, and I’m down books row G.
If you’d rather check every figure first, the same quote waits on a signed-in page with a single Buy button.
When a buy takes more than one approval, Nook collects each person's approval for their share in the chat apps your people already live in.
Family chat3 of 3 approvedYour agent sends one connect link, gets a real quote back from the store’s own checkout, collects the shopper’s approval in chat and holds a confirmed order number before the first minute is up, all through four calls to the same small API.
The shopper opens your one link and connects their stores.
Nook returns the store’s real total, line by line.
The shopper confirms the total in their own words.
The store’s order number comes back on the receipt.
Illustrative timings, with the quote step set to the 19 s median shown on this page rather than measured end to end.
# 0:08 connect
POST /v3/api/agent/connect-links
→ { url }
# 0:27 quote
POST /v3/api/agent/quotes → 202
GET /v3/api/agent/quotes/q_…
→ ready, total_cents: 2349
# 0:39 approve
POST /v3/api/agent/purchases
{ quote_id, confirmation: "approved" }
# 0:53 order confirmed
GET /v3/api/agent/purchases/p_…
→ confirmed, receipt.order_idNook runs every store action as one of its own fixed, guarded steps and never runs browser code that a model wrote, which is a deliberate design choice that keeps both your orders and your data out of a model’s reach.
Every store action runs as one of Nook’s fixed steps, and the step that presses a store’s final order button only does so after the shopper has approved that exact quote.
An agent can ask Nook for a quote or a purchase, but it can never hand Nook code or requests of its own to run, so nothing a model writes can place an order or read your session cookies on the side.
Raw store responses are full of personal data and session tokens, so they stay inside Nook and your agent receives clean quote data with the items, the charge lines and the total.
Nook speaks REST, MCP and ChatGPT plugins, so you can bring your own model and keys and turn a person’s store connection into a checkout your agent can actually finish.
Self-serve partner keys (nk_…) are stored as hashes only, and one small REST API covers quotes, purchases, preferences and order status, idempotent end to end.
Nook runs as an MCP server, so connect, quote and purchase become tools that any MCP client, including Claude, Cursor or your own runtime, can call.
You can ship a GPT or plugin that actually shops, because the people using it connect their accounts once and your GPT checks out through Nook.
Muse, GPT, Grok and open-weights models can all finish a checkout as long as they can call a tool, because the protocol doesn't care which model is driving.
Business accounts hold partner keys per agent, product or team, while each person's connections and consent stay theirs, so your keys never become their credentials.
Fine-tune a model on your own store's flows: real sign-ins, carts and checkouts as the training signal, so your agent learns your store's quirks instead of guessing.
Start with the guide at nookstore.xyz/agents.md, or go deeper in the docs.
Nook quotes delivery the way each store’s own checkout does, with the real fee, the delivery window and your saved address, so groceries, takeout, pet food and household restocks arrive when the quote said they would.
Any agent can register a partner key and shop through Nook, including the chat apps and agent platforms yours already lives in.
Nook runs on a production Plaid account, so next to your stores your agent can pay from the bank you choose, whether that’s Chase, Wells Fargo, Bank of America or one of thousands more institutions, while you sign in at your bank through Plaid and Nook never sees the password.
A single phone number can connect to many bank accounts through Plaid, so every bank you link sits behind the same sign-in as your stores.
Wells Fargo
Bank of America
Citi
Capital One
U.S. Bank
PNC
Truist
TD Bank
Ally
Mercury
Brex
Here are the short answers, and the full guide lives at nookstore.xyz/agents.md.
Nook connects your store accounts so your agent can shop for you, preparing real checkout quotes from the stores you already use and buying only once you approve.
Amazon, Uber Eats, Instacart, Walmart, Target and Chewy can quote and buy, while DoorDash, Urban Outfitters, Resy, Detoure, Delta and Apple connect only for now.
Your agent sends you a Nook link, and you sign in with your phone, connect your stores and go back to your agent, while agents get a partner key by registering with the guide at nookstore.xyz/agents.md.
Nook is in private beta without public pricing yet, so connecting your accounts is free during the beta and agents register their own partner key for free.
No, because every purchase needs your approval, either in your own words replying to the quote's total in chat or with Buy on Nook's review page, and if the total comes in higher than the quote the purchase stops instead.
Saved store sessions are encrypted, and your agent gets an authorized connection to shop for you rather than your password or your browser session.
Four steps between your agent and a real order.
Your agent sends you one link, and you sign in to Nook with your phone and connect your store accounts in one tap while your agent never sees what you type.
Nook works through your signed-in account, finds the product, and reads the store's own checkout figures.
You see the product, quantity, price, address and payment before anything is bought, and saved preferences fill only what a request leaves out.
You approve in your own words in chat or with Buy on Nook's review page, and Nook presses the store's final order button once and reports the order number back.