AI Agents WATCH CLOSELY OPPORTUNITY

Anthropic's free starter kit for AI shop assistants

ShiniFrontiers · Source: anthropics

Section 1 of 6 FREE PREVIEW

What problem does it solve?

Commerce Agents is free starter code from Anthropic for two separate AI assistants built on Claude: one that helps a store's customers shop, and one that helps its staff run the back office. It is a blueprint: the store's developers connect it to the store's own systems.

A. Shopping assistant — for the store's customers

Problem: Customers can't find the right product, or wait for staff to answer simple questions.

What it does: Searches the store's catalogue, compares products, fills the cart, and answers order and returns questions.

B. Merchant assistant — for the store's staff

Problem: Staff spend hours on reports, product listings, stock alerts and promotions.

What it does: Explains sales, and drafts listing fixes, restocks, price changes and campaigns.

Note: Every change waits for a person to approve it.

See the difference

Who asks	Today	With Commerce Agents
Shopping assistantA customer: "A gift for my dad, who walks every morning, under $60"	Filters can't search for that, so the customer browses or asks staff	Gets three suitable walking shoes, each with a reason, ready to add to the cart
Shopping assistantA customer: "Can I return my shoes?"	Waits for staff to look up the order and the returns policy	Gets an answer straight from the store's order records and returns policy
Merchant assistantA manager: "Why did sales drop last week?"	An analyst pulls reports and writes a summary	Gets a plain-English explanation from the store's sales figures
Merchant assistantStock of a best-seller runs low	A staff member spots the alert and places the reorder	The assistant drafts the reorder; a staff member approves it

Illustrative examples.

What it saves: Staff time on routine customer questions and store admin.
What it costs: The code is free. Claude usage is billed per conversation: by Anthropic, or by AWS, Google Cloud or Microsoft if the store runs it through them.
How ready: Proven in the lab

Working demo code for four example industries, with made-up companies. A blueprint, not a product: the store's developers connect and secure it.

Join free to unlock: how it works, a step-by-step setup with its costs, what you gain, and where it falls short, for every Friday idea.

Join free to read it all →

Section 2 of 6 MEMBERS

How does it work?

Join free to read this section

Anthropic's free starter kit for AI shop assistants — how it fits together.

Join free to unlock: how it works, a step-by-step setup with its costs, what you gain, and where it falls short, for every Friday idea.

The big picture

Picture two new staff who work only from the store's own records. The shop assistant can fill a customer's basket but never takes payment; the till does that. The back-office clerk can fill in order forms and price changes, but nothing happens until the manager signs.

How the two assistants fit into a store's systems

A customer's or staff member's request travels down the stack. Green is new; the dashed line is where chat text goes out to the AI.

Front end. Storefront: where customers shop and chat. Staff portal: where staff ask questions and approve changes. Both assistants plug into what the store already runs; checkout is unchanged.

Assistants. Shopping assistant (A): searches products, compares them, fills the cart, answers order and returns questions. Merchant assistant (B): explains sales; drafts reorders, listing fixes, price changes and campaigns. Ten shop tasks ready, five each. Products and prices shown are checked in code against the store's records.

Connections. Store connections: the store's developers write each one: catalogue, cart, orders, policies, sales and stock. The connection points are defined; the developers build and secure them.

Store systems. Catalogue, orders, stock, sales: the store's own records; the assistants read only these. Existing checkout: takes payment; the AI never sees payment details. No new payment risk, and no price or stock change without a person's approval.

A. Shopping assistant: technical view

B. Merchant assistant: technical view

Why not just connect an AI to the store?

A developer can connect an AI such as ChatGPT or Claude straight to a website. It is smart, but it arrives with no knowledge of the store, no access to its systems and no safety checks. Commerce Agents comes with the checks built.

	Plain AI connection	Commerce Agents
Connecting the store's systems	❌ Developers work out what is needed and build it all	⚠️ A defined set of connection points; developers write each one
Products and prices on screen	❌ Whatever the AI writes	✅ Checked in code against the store's records
Payment	❌ Developers design a safe way	✅ Handed to the store's existing checkout; the AI never sees payment details
Changes to prices, stock, listings	❌ Could act on its own unless blocked	✅ Drafted only; a staff member approves each one
Limits on changes	❌ Must be built	✅ Caps on price moves, discounts, restock size and campaign budget, checked when drafted and again when approved
Text hidden in product data or messages	❌ The AI may follow it (prompt injection)	✅ Fenced off as material to read, not instructions to obey
Shop tasks	❌ Each written from scratch	✅ Ten ready: five for customers, five for staff

What the code enforces holds whatever the AI says. What is only asked of the AI, such as quoting figures only from the store's data, can still slip in its written reply; that is a wrong sentence to correct, never an order, payment or change to reverse.

A. Shopping assistant: helps the store's customers

Who	What happens	In everyday terms
1. The customer	Types "a gift for my dad, who walks every morning, under $60"	Asks a shop assistant for help
2. The assistant	Searches the store's product list, prices and stock	Checks the shelves and the stock room
3. The assistant	Shows three suitable walking shoes, each with a reason	Brings three pairs to try
4. The customer	Adds a pair to the cart and pays on the store's normal checkout page	Pays at the till, not to the assistant

The assistant can remember what a customer said before, such as shoe size or budget; that memory is personal data the store must manage.

B. Merchant assistant: helps the store's staff

Who	What happens	In everyday terms
1. A staff member	Asks "Why did sales drop last week?", or the assistant notices stock is low	Asks the back-office clerk
2. The assistant	Reads the store's sales figures, listings, stock and prices	Pulls the files
3. The assistant	Explains what happened, or drafts a change: a reorder, listing fix, price change or campaign	Fills in the form
4. A staff member	Approves or rejects the change; nothing changes until someone approves it	The manager signs, or doesn't

Both assistants

Run on Claude, Anthropic's AI: through Anthropic by default, or through AWS, Google Cloud, Microsoft or the company's own AI gateway.
Read only the store's own systems, through connections the store's developers build; every product, price and figure shown is checked against those records, which is grounding in practice.

Join free → I have an account

Section 3 of 6 MEMBERS

How to set it up, and what it costs

Join free to read this section

Anthropic's free starter kit for AI shop assistants — steps and cost.

Join free to unlock: how it works, a step-by-step setup with its costs, what you gain, and where it falls short, for every Friday idea.

Before you start: is it a fit?

	A. Shopping assistant	B. Merchant assistant
Good fit if	Many customers ask the same questions: delivery, returns, "which one should I buy?"	Staff spend hours each week on reports, listings or stock
	Customers search in ways filters can't handle	Sales questions wait days for an analyst
Who acts on it	E-commerce or customer service head, with a developer	Operations manager or store owner, with a developer
Not yet if	Few products and few questions; filters and staff already cope	Sales data is spread across tools nobody has connected

Before you begin

You need	Why
A developer	To install it, connect it to the store's systems and secure it
An Anthropic API key	The demo's default way to use Claude; AWS, Google Cloud or Microsoft also work later
Python 3.11 or later, and Node 22	The programming tools the code needs; the developer installs them

Part 1: Try the demo (both assistants)

Download and install (the developer)
Start the demo shop and staff screen (the developer) Expect: the demo shop at localhost:3000 and the staff screen at localhost:3100.
Try both assistants (the store owner) On the shop page ask: "A gift for my dad, who walks every morning, under $60." On the staff screen ask: "Why did sales drop last week?" Travel, telecom and ticketing demos also ship. Expect: product suggestions with reasons, and a plain-English sales explanation, all on made-up demo data.

Part 2: Build a starter version for the store

Let Claude Code scaffold it (the developer) Anthropic's coding tool asks about the store's systems, plays the plan back, and builds the project on these packages.
Start small (the developer) A shopping pilot connects only search and product details. A merchant pilot connects only the read-only data, with every change refused. Switch off anything the store doesn't have, such as order tracking.
Test with real questions (the customer service or operations manager) Ask the questions customers and staff really ask, and check every answer against what you know is true.
Measure the AI cost (the developer) Every AI call logs its usage. Run a sample of real conversations and read the logs; that, not a guess, is the cost per conversation.

Part 3: What the store must add before customers see it

Who	What to add	Why
The developer	Login and access checks on every page and connection	The demo accepts anyone
The developer	Abuse limits on the chat	To stop misuse and runaway costs
The developer	The store's own business rules: fraud, eligibility, pricing, stock	The code checks limits; the store decides what is allowed
The operations manager	Real limits on price moves, discounts, restocks and budgets	The shipped limits are demonstration values
The operations manager	Who approves changes, and a daily check of the queue	Nothing happens until someone approves
The data or privacy lead	How long customer memory is kept, and how customers see and delete it	What the assistant remembers is personal data

What it costs

Item	Cost
The code	Free (Apache 2.0 open-source licence)
The AI it runs on (Claude)	A usage fee per conversation, paid to Anthropic, or to AWS, Google Cloud or Microsoft. The project gives no figures; measure it in the pilot (Part 2, step 4).
People	Developer time to connect, secure and maintain it, since Anthropic does not maintain it
Hardware	None new. It runs on an ordinary web server; the AI runs at the provider.

Join free → I have an account

	What changes	What it can mean for the store
A	Questions answered at any hour, from the store's own data	More sales kept, fewer routine questions reaching staff
A	Payment stays on the store's existing checkout	No new payment risk to manage
B	Sales explained; reorders and listing fixes drafted	Staff time moves from admin to decisions
B	A person approves every change, within set limits	Speed without surprise price or stock changes

Limit	Why it matters	What to do
Not maintained by Anthropic	It is a one-time blueprint: no updates, no fixes, no outside contributions.	Plan for the store's developers to own and update the code.
Not safe to put online as it is	The demo has no login, no abuse limits and placeholder change limits.	Add everything in Part 3 before any customer sees it.
No ready connections to shop platforms	It does not plug straight into the store's shop software, payments or analytics.	Developers connect each system; the guide names common ones (Stripe, PayPal, QuickBooks, Snowflake, BigQuery).
Runs on one AI (Claude)	Every conversation has a usage fee, and the store depends on Claude.	Measure cost in the pilot; billing can go through Anthropic, AWS, Google Cloud, Microsoft or a gateway.
Written replies can misstate a figure	Shown products and prices are checked, but a sentence can still be wrong.	Test answers before launch; a misstatement can never trigger an action.
Memory works in one setup only	Of the three ways to run it, only one includes automatic memory.	If memory matters, choose that setup when building.
Only as good as the store's data	Thin product details mean weak suggestions; unconnected sales data is invisible to it.	Check product descriptions, and connect the main sales system first.

Anthropic's free starter kit for AI shop assistants

What problem does it solve?

How does it work?

The big picture

How the two assistants fit into a store's systems

Why not just connect an AI to the store?

A. Shopping assistant: helps the store's customers

B. Merchant assistant: helps the store's staff

Both assistants

How to set it up, and what it costs

Before you start: is it a fit?

Before you begin

Part 1: Try the demo (both assistants)

Part 2: Build a starter version for the store

Part 3: What the store must add before customers see it

What it costs

What do you gain?

What are the limits?

What this starter kit can't do yet

What it does not do

Sources