# Big Brands, AI Agents, and What a Small Shop Can Copy What DoorDash, Amazon, Starbucks and Chipotle actually offer, what a small business can copy, and two practical kits for catalogs, order links and staff help. Author: J.A. Watte Published: November 7, 2026 Source: https://jwatte.com/blog/big-brands-mcp-agentic-commerce-small-business/ --- **Let the assistant help people discover, compare and choose; keep the order, account and payment in your business's established checkout process. Publish useful menu, product and store facts without a customer account, and require sign-in when private records or account actions need it. This week, a small shop can clean up those facts and make one useful link to its existing order or booking page. The big-brand examples below show both this pattern and the exceptions.** Starbucks makes the idea unusually clear: its [ChatGPT listing](https://chatgpt.com/plugins/plugin_asdk_app_69cf433334cc8191951af77fbe29f96c) describes drink recommendations, finding a store, and placing an order through the Starbucks app or website. The assistant helps with the choice; the customer reaches the brand's ordering process. We observed that listing on November 7, 2026 at 18:54:26 UTC. That is a useful starting point for a cafe, retailer, service business or group of locations. A cart, an inventory hold and a payment have different consequences. Some large platforms support purchases inside chat. We recommend an external checkout as the first design for a small business that already has a working order page. ## First, separate the names from the work An **AI agent** is software that can use tools to carry out parts of a task, rather than only write an answer. A **tool** is a named function, such as finding stores or reading today's menu. **Model Context Protocol**, or **MCP**, gives compatible assistants a common way to discover and call tools. An **MCP server** offers those functions; a **connector** is the connection the customer enables in an assistant. The [MCP documentation](https://modelcontextprotocol.io/docs/learn/architecture) explains those roles. An assistant **app** or **plugin** is the product experience exposed inside that assistant. We use the names each listing displays. A **directory** helps people find those experiences. A company can have an agent without publishing an MCP server, and an official MCP server can exist without appearing in the directory search you tried. Those distinctions matter in the DoorDash and Chipotle examples. For an owner, the practical question is what the connection can read, change or send. A public menu needs different access from a customer's saved address. A recommendation needs different permission from charging a card. Keep those questions visible even when the product name sounds familiar. ## The pattern visible in the listings We checked these public listing descriptions in a signed-in browser. This establishes what the directory said at the stated time. We did not connect the brands' tools, create carts or place orders. | Listing | What the assistant helps with | Account boundary | What a small business can copy | |---|---|---|---| | [Starbucks in ChatGPT](https://chatgpt.com/plugins/plugin_asdk_app_69cf433334cc8191951af77fbe29f96c) | Drink recommendations and store finding | Ordering continues through its app or website; sign-in behavior was not tested | Suggest published choices and open the right order page | | [Domino's India in Claude](https://claude.ai/directory/dominos-india) | Menu browsing, store lookup and further ordering tools | Listing says sign-in not required; account tools also exist | Keep public discovery distinct from loyalty, addresses and order history | | [Instacart in Claude](https://claude.ai/directory/instacart) | Product search and cart help | Listing says sign-in required | Let the person build a useful shopping selection within supported account permissions | | [Uber Eats in Claude](https://claude.ai/directory/uber-eats) | Nearby restaurants and dishes | Listing says sign-in not required | Answer a useful local question before asking for an account | The Domino's India, Instacart and Uber Eats observations were at 18:53:15, 18:53:10 and 18:53:36 UTC respectively on November 7, 2026. Domino's India is made by Jubilant Foodworks Limited and applies to India. Its tool list includes account-related functions; the listing's overall sign-in label must not be read as permission to retrieve anyone's saved addresses. The Uber Eats listing showed search and an analytics tool, so even discovery needs a data-handling explanation. Those details come from the linked listings. Anthropic's [April 23, 2026 connector announcement](https://claude.com/resources/articles/connectors-for-everyday-life) describes recommending connected services and asking before bookings or purchases. That is a platform design statement, separate from our observations of individual listings. It does not establish that every merchant uses the same checkout screen.
Public business facts lead to assistant recommendations, then the business checkout where the customer reviews and pays; private staff tools stay behind sign-in.
The big-brand pattern, small-business sized: a recommended first implementation. Some platforms also offer approved checkout inside chat.
```text The big-brand pattern, small-business sized A practical starting point for your business 1. Public facts Menu, catalog, hours and stores ▼ 2. Assistant helps choose Answers questions and recommends ▼ 3. Link to your own page Checkout or booking, choices filled in ▼ 4. Customer reviews and pays Sign-in, when needed, happens there ▼ 5. Staff confirm when needed Availability, exceptions or final approval Separate staff lane: Staff-facing agent → hiring, inventory and seller tools → behind sign-in, with human approval ``` ## DoorDash: grocery handoff, corporate MCP and merchant help DoorDash's [December 17, 2025 announcement](https://ir.doordash.com/news/news-details/2025/DoorDash-Launches-Grocery-Shopping-App-Within-ChatGPT/default.aspx) describes a ChatGPT grocery app: connect a DoorDash account, turn recipe ingredients into a shopping list, then open the cart in DoorDash to review, adjust and buy. The announcement began with select users. Our November 7 directory check found the listing, but did not test individual account or regional availability. There is also a different, official MCP product. DoorDash [announced its corporate ordering connector on September 30, 2026](https://about.doordash.com/en-us/news/doordash-ordering-connector). Its [developer guide](https://developer.doordash.com/en-US/docs/mcp/tutorials/get_started/) labels access private beta for approved testers. Employees sign in to their own DoorDash accounts; a company's agent can find restaurants, build a cart and place an order. The guide tells developers to call the order tool only after the person explicitly confirms, because it charges that person's payment method. This is substantially more capable than a public menu feed. The distinction corrects a tempting conclusion: we found no DoorDash connector in our [Claude directory search](https://claude.ai/directory?q=doordash) during the November 7, 18:52:25 to 18:54:48 UTC check. That does not mean DoorDash has no MCP server. Company documentation establishes the separate beta. For merchants, DoorDash's [April 9, 2025 tools announcement](https://about.doordash.com/en-us/news/doordash-unveils-ai-powered-tools-to-enhance-online-menus-and-streamline-merchant-operations) describes item-description and photo tools in its business products. Its [September 30, 2026 DashBuddy announcement](https://about.doordash.com/en-us/news/dashbuddy) describes onboarding help for Dashers. These sources document operational AI uses; sales attributable to MCP remain unestablished. DoorDash's [second-quarter 2026 shareholder letter](https://ir.doordash.com/news/news-details/2026/DoorDash-Releases-Second-Quarter-2026-Financial-Results/default.aspx) also discusses Ask, its restaurant-discovery and grocery-basket assistant, alongside merchant onboarding and catalog processing. The AI discussion does not isolate revenue from those features. Companywide growth would be the wrong measure to present as agent-generated sales. **What a small business can copy:** make the menu accurate and connect recommendations to the existing order route. For staff, answer recurring onboarding questions from an approved reference, with a manager available when the answer is unclear. Copy the clear task and approval boundary before considering an order-submission tool. ## Amazon: distinguish the seller, builder and shopper products Amazon's [Selling Partner connector](https://claude.com/marketplace/connectors/amazon-selling-partner) is a seller-facing beta. The listing requires sign-in, uses Seller Central data for sales, inventory and listing questions, and describes review and approval of actions. It includes both reading and writing tools. Its purpose is private business assistance. AWS, Amazon's cloud platform, has a separate [managed MCP Server, generally available since May 6, 2026](https://aws.amazon.com/about-aws/whats-new/2026/05/aws-mcp-server/), for coding agents working with AWS documentation and services. [Bedrock AgentCore became generally available on October 13, 2025](https://aws.amazon.com/about-aws/whats-new/2025/10/amazon-bedrock-agentcore-available/) as infrastructure for operating agents, including identity controls and connections to tools. A small retailer does not need to adopt that infrastructure merely to publish product facts. For shoppers, Amazon's [shopping overview](https://www.aboutamazon.com/news/retail/amazon-agentic-ai-gen-ai-shopping) notes that Rufus was renamed **Alexa for Shopping on May 13, 2026**. It describes recommendations using catalog details and shopping context. [Alexa+](https://www.aboutamazon.com/news/devices/new-alexa-generative-artificial-intelligence) is the broader assistant, with integrated services for tasks such as reservations and food ordering; the cited announcement was updated July 21, 2026. Product names and availability should not be treated as interchangeable. The most directly useful merchant example is the [Shop Direct Merchant Workspace announced November 4, 2026](https://www.aboutamazon.com/news/retail/amazon-shop-direct-merchant-workspace). Eligible merchants verify their websites and provide catalog feeds, including through supported feed providers. Customers can discover an item on Amazon and follow Shop Direct to the merchant's site to purchase. The merchant retains responsibility for pricing, inventory, fulfillment and support. This documented feed program is not itself proof of an MCP server. Amazon also identifies an exception: eligible **Buy for Me** items let customers confirm order details on Amazon, after which Amazon completes the purchase on the merchant's website. Its [original April 3, 2025 explanation](https://www.aboutamazon.com/news/retail/amazon-shopping-app-buy-for-me-brands) describes the transfer of required checkout details. That authorized transaction system has responsibilities beyond preparing a link. Keep personal and payment data out of URL parameters. Amazon's [November 4, 2025 statement about outside shopping agents](https://www.aboutamazon.com/news/company-news/amazon-perplexity-comet-statement) says agents that buy on a customer's behalf should operate openly and respect a business's decision whether to participate. That source establishes Amazon's position on that date. Any later legal steps and outside agents' current access are beyond the evidence reviewed here. **What a small business can copy:** maintain one reliable catalog, consider official distribution programs your provider supports, and keep private sales analysis behind staff access. Have a person review proposed listing or stock changes. The catalog can serve discovery channels without giving those channels unrestricted control of the business. ## Starbucks: recommend the drink and help the barista The [Starbucks ChatGPT listing](https://chatgpt.com/plugins/plugin_asdk_app_69cf433334cc8191951af77fbe29f96c) identifies Starbucks Coffee Company as maker and describes preferences leading to drink suggestions, store finding and the Starbucks app or website for ordering. Its [April 15, 2026 beta announcement](https://about.starbucks.com/stories/2026/meet-the-beta-starbucks-app-in-chatgpt-a-new-way-to-discover-your-next-favorite-drink/) explains suggestions from words or a photo, customization, and choosing a coffeehouse before that handoff. Starbucks' [AI update, revised June 2026](https://about.starbucks.com/press/2026/supporting-the-moments-that-matter-with-artificial-intelligence/), also describes the app as a beta. We did not test its account flow or a purchase. Staff assistance is further along. [Starbucks announced Green Dot Assist on June 10, 2025](https://about.starbucks.com/press/2025/meet-green-dot-assist-starbucks-generative-ai-powered-coffeehouse-companion/) as a pilot in 35 coffeehouses, answering barista questions on store iPads. Starbucks' [January 28, 2026 earnings transcript, page 5](https://s203.q4cdn.com/326826266/files/doc_financials/2026/q1/SBUX-1Q26-CORRECTED-TRANSCRIPT.pdf) says it fully scaled across North American coffeehouses in November 2025, supporting beverage instructions and operational questions. Calling it only an early pilot would now miss that company-reported rollout. **What a small business can copy:** give customers recommendations tied to actual menu choices, and give staff a separate helper for approved recipes, equipment instructions and routines. Keep source dates visible and refer uncertain answers to a supervisor. Neither source isolates sales caused by the assistant. ## Chipotle: hiring help counts, even without a shopping plugin Chipotle's [October 22, 2024 Ava Cado announcement](https://newsroom.chipotle.com/2024-10-22-CHIPOTLE-INTRODUCES-NEW-AI-HIRING-PLATFORM-TO-SUPPORT-ITS-ACCELERATED-GROWTH) describes a conversational hiring assistant from its hiring software partner, Paradox. It answers candidate questions, gathers information and schedules interviews; managers choose candidates for offers. This is an administrative agent. The source does not document a public Chipotle MCP server. Its [February 19, 2025 follow-up](https://newsroom.chipotle.com/2025-02-19-CHIPOTLE-AIMS-TO-HIRE-FUTURE-BILLION-DOLLAR-BUSINESS-LEADERS-IN-NEW-BURRITO-SEASON-HIRING-PUSH) reports average application-to-start time falling from 12 days to four. That is a company-reported operational result. It is different from the anticipated percentage improvement in the launch announcement, and it does not establish additional restaurant revenue. Chipotle separately [announced in-restaurant tests of Autocado and an augmented makeline on September 16, 2024](https://newsroom.chipotle.com/2024-09-16-CHIPOTLE-DEBUTS-AUTOCADO-AND-THE-AUGMENTED-MAKELINE-BY-HYPHEN-IN-RESTAURANTS). Kitchen machinery is another category; that source does not establish chainwide deployment or MCP use. Our [ChatGPT search](https://chatgpt.com/plugins?search=chipotle) found no Chipotle plugin during the November 7 browser check. Neither observation means the company has no AI program. **What a small business can copy:** begin with an approved job FAQ and interview scheduling assistance, while the hiring manager makes employment decisions. The sources do not establish the candidate sign-in flow, so we do not invent one. For your implementation, keep applicant records protected and separate from the public customer helper. ## Other examples worth copying carefully ### Instacart, Uber Eats and Domino's India: reduce the work of choosing The [Instacart listing](https://claude.ai/directory/instacart) connects meal ideas to product search and a cart. The [Uber Eats listing](https://claude.ai/directory/uber-eats) connects a local food question to restaurants and dishes. [Domino's India](https://claude.ai/directory/dominos-india) exposes both discovery and deeper ordering functions. These are distinct products with different account requirements, as the table above shows. **What a small business can copy:** help a customer get from a need to a specific choice. A cafe can map published drink options to preferences; a garden shop can identify which in-stock product fits a stated requirement. Preserve the selected item and location when sending the person to the next page. Never turn an uncertain dietary answer into an assurance that an item is safe. There is a useful platform comparison here. [Instacart's help page](https://www.instacart.com/help/section/2893565984/349218736) says Claude users build a cart, then select Checkout on Instacart to confirm details and pay on Instacart. By contrast, the [December 8, 2025 OpenAI announcement](https://openai.com/index/instacart-partnership/) describes Instacart's checkout inside ChatGPT. Both involve account connection and customer review. The merchant's authorized transaction process matters more than whether the final screen appears in a separate browser tab. ### Walmart: turn a task into a relevant basket Walmart's [October 30, 2026 Sparky update](https://corporate.walmart.com/news/2026/10/sparky-shopping-assistant-expansion) describes photo-based product help, turning a list into a cart, recipe assistance and finding items inside a store. It also distinguishes availability from plans: voice capabilities were expected chainwide by the end of the year. The announcement describes features without isolating their effect on sales. **What a small business can copy:** organize products around real tasks, such as the supplies for a small repair or a published catering menu for a meeting. Keep each suggestion tied to a real item and location. Let the customer review substitutions before anything is ordered. ### Shopify: use today's interface, not last year's tutorial A **catalog feed** is a structured list of products, variants and availability. Shopify's current [Storefront MCP migration page](https://shopify.dev/docs/apps/build/storefront-mcp) says the old catalog and cart tools were removed. It directs developers to **Universal Commerce Protocol**, or **UCP**, a set of commerce interfaces covering catalog, cart and checkout work. It also points merchants seeking a built-in chat experience to Shopify Inbox's agent, subject to its requirements. **What a small business can copy:** if Shopify already holds your products and orders, ask your web person to evaluate its supported tools before commissioning another inventory database. Keep the product source authoritative and check current documentation before following an older MCP example. Our earlier [industry tour](/blog/blog-mcp-servers-by-industry-guide/) is useful background; for implementation details, follow Shopify's current documentation. ### Square: a private merchant connection needs narrower permissions Square's [MCP documentation](https://developer.squareup.com/docs/mcp) describes a beta connection to its merchant APIs, the software interfaces for customers, orders, items and more. It recommends **OAuth**, a sign-in process that grants a service specific permissions without handing the assistant your password. Square says its remote service accesses production resources and recommends testing in a seller's Sandbox, a separate test environment, first. **What a small business can copy:** let an authorized manager ask bounded questions about the records already in the business system. Do not expose a broad merchant connection as a public menu tool. Read the actual requested permissions; a familiar brand does not make every available action appropriate. ### Toast: begin with the system your restaurant already uses Toast's [October 29, 2025 Toast IQ announcement](https://pos.toasttab.com/news/toast-expands-toast-iq-smart-ai-assistant) describes an assistant for operators using business data to answer questions and take actions, including menu updates and shift edits. The launch described availability to US customers through Toast Web and Toast Now. It is a staff-facing use of the restaurant's business records. **What a small business can copy:** evaluate the reporting and assistance already in your point-of-sale system, the system recording your sales. Start with questions about approved data and review proposed edits. We recommend that review boundary; the announcement reviewed leaves its approval controls unspecified. ## How a small or medium business adapts this ### 1. Pick the customer decision before the technology **Who does this: owner and the person who handles customers.** Choose a repeated question that ends in a useful next step. Examples: which location is open, which published menu choice fits a preference, which product variant is available, or which service request form fits the job. Write down what currently happens when staff miss the question. Does the customer leave, call again, choose the wrong location, or submit a request without enough information? That gives the project something observable to improve. **Done when:** you can describe the task, its source of truth and its finish line in ordinary language. "More AI visibility" is too vague to serve as the only success measure. ### 2. Prepare the facts for your business type **Who does this: operations lead and web person.** A stable ID is a label that keeps identifying the same item or location even if its display name changes. Use those labels to keep the answer and the order page connected. | Business | Public facts to prepare | Useful handoff | What stays with the person or staff | |---|---|---|---| | Restaurant or cafe | Menu, published prices and allergens, service hours, locations | Existing online ordering with item and store choices | Allergy questions beyond published facts, substitutions and purchase confirmation | | Retail or ecommerce | Product and variant IDs, availability, fulfillment and return terms | Product page or existing checkout | Final stock and total, account details and payment | | Service business | Service scope, area, exclusions and quote basis | Relevant quote or booking form | Private job details, agreement and staff acceptance | | Multiple locations | Location IDs, addresses, local hours and holiday exceptions | The selected location's action page | Clarifying ambiguous locations and confirming local capacity | Include when changing facts were checked. If the inventory source is unavailable, say availability cannot be checked. Do not translate that failure into "sold out," and do not present yesterday's stock as current. For food, copy the business's approved allergen statements accurately. Missing information is unknown. A menu photo or a product name is not sufficient evidence for an allergy claim. The kit provides separate fields for published information, uncertainty and the staff contact route. **Done when:** a staff member can compare the feed with the customer page and explain every difference. Assign one person to correct hours, prices and product changes across both. ### 3. Build a useful handoff before adding transactions **Who does this: web person or developer.** A handoff link opens the existing page with approved choices selected. A **query parameter** is a named choice in that address, such as a public item ID. It is not a suitable place for a customer's name, phone number, address, health information or payment details. The destination must actually understand those choices. A link with a plausible item parameter is useless if the checkout ignores it. Keep the destination fixed to an approved business page, reject unknown fields, and retrieve the current price there. Opening the link must not reserve stock, send an email or place an order. The companion kit uses `build_order_link` to construct a URL only. Its **read-only annotation**, a label describing that the tool makes no change, is appropriate only while the implementation stays within that boundary. A tool that saves a cart or holds an appointment needs different treatment, even if no payment occurs. **Done when:** the person reaches the correct choice, sees what still needs confirmation, and can continue through the normal form. Test the same page without prefilled choices so the ordinary customer route still works. ### 4. Keep staff assistance behind staff access **Who does this: owner and whoever manages accounts.** Start with approved procedures or an authorized inventory summary. Give the helper only the material required for that task. Ask it to name the source and revision date, and to send uncertainty to a person. A small cafe's first staff assistant might answer preparation questions from its own approved manual. A retailer might use one to flag inconsistent product descriptions. A service company might draft an interview schedule from a manager's approved availability. These are our proposed small-team adaptations. **Done when:** the owner can list allowed records, prohibited actions and the person responsible for reviewing output. Public menu access must not become access to applicant files or customer order history. ### 5. Test the task, then consider distribution **Who does this: developer, then owner.** Try normal inputs and the cases that usually break: unknown items, closed locations, stale inventory, missing allergen information, an altered destination, and a repeated request. A failed lookup needs an honest explanation and a usable customer route. For construction, use [our connector guide for businesses, nonprofits, events and WordPress](/blog/build-mcp-connector-nonprofits-events-wordpress/). For getting the finished service found, use [the directory listing guide](/blog/mcp-directories-listing-guide-2026/). A **registry** records server descriptions and addresses; a directory listing is a separate distribution and review process. Neither is a substitute for completing the actual customer task. **Done when:** owner-approved testing shows the supported task works, documentation matches its behavior, and privacy and support information are ready. Directory publication still depends on the platform's rules and review. ## Small-business proof from our own connectors In our case, these patterns are already concrete. We build connectors for a few small businesses of our own, shown here under fictional names: Kestrel Lodging, a twelve-property motel and extended-stay group; Quarry Road Storage, a self-storage facility; and Brookfield Laundry and Millrace Laundry, two unattended laundromats. Our November 7, 2026 read-only checks found working public information tools on all four connectors. The checks were at 18:23:47 UTC, followed by connector reads at 18:24 UTC. They establish working information paths; resulting sales were not measured. Kestrel Lodging offers property information, availability and stay quotes, with a prefilled booking-form route. Quarry Road Storage offers unit information, hours and an availability answer with a checked time. The laundromat tools answer practical questions about machines, services and hours. Millrace Laundry's contact-link tool prepares an inquiry link; the person adds contact details, reviews it and sends it. The [form handoff article](/blog/lodging-sites-ai-agents-booking/) explains the general pattern. Be precise about the boundary: Kestrel Lodging also has separate booking actions that can send verification email and hold inventory. Its whole server is not read-only. Those actions were excluded from this research. The [direct-booking article and kit](/blog/mcp-connector-listing-small-business/) document that deeper implementation and its staff confirmation requirements. Our starter kits here deliberately keep customer details out of links and limit the assistant to public facts and preparation. These sites show that a small operator can expose useful information without first building a large marketplace. They do not establish that assistants delivered customers or increased revenue. ## Example connector specs you can adapt These examples adapt public patterns and ideas using our own wording and tools; we copied no company code or listing wording, and no affiliation is implied. - [Restaurant ordering connector example (.md)](/downloads/restaurant-ordering-connector-example.md): public menus, locations, hours, published allergens and dated availability, with an order link and an optional account-history boundary. - [Grocery and retail cart handoff example (.md)](/downloads/grocery-and-retail-cart-handoff-example.md): product search, stock checks and a cart review link, plus the separate decision about platform checkout programs. - [Recommend and find a store example (.md)](/downloads/recommend-and-find-a-store-example.md): suggestions from your published catalog, allowed customizations and a route to the selected location's checkout. - [Seller and owner data connector example (.md)](/downloads/seller-and-owner-data-connector-example.md): private reports, inventory and proposed price or listing edits that require owner approval in your business system. - [Operator assistant connector example (.md)](/downloads/operator-assistant-connector-example.md): staff questions answered from your records and manual, with proposed menu or shift changes for manager review. Each file includes original tool definitions, fictional request and response examples, data boundaries, privacy wording and a test plan. Give your developer the closest example alongside the readiness and catalog kits below. Replace the sample facts, enforce the stated limits and test your own checkout or staff system before offering the connector to anyone. ## Where the revenue comes from Amazon provides a concrete but bounded signal. In its [November 18, 2025 shopping article](https://www.aboutamazon.com/news/retail/amazon-agentic-ai-gen-ai-shopping), the company reported that customers who use Rufus while they shop are "60%+ more likely to convert." That is an association reported by Amazon. It does not prove the assistant caused a 60% increase, and it is not a forecast for an independent store. The [DoorDash app announcement](https://ir.doordash.com/news/news-details/2025/DoorDash-Launches-Grocery-Shopping-App-Within-ChatGPT/default.aspx) presents recipe-to-groceries discovery as a new way for local merchants to reach customers. [Amazon's Shop Direct announcement](https://www.aboutamazon.com/news/retail/amazon-shop-direct-merchant-workspace) describes product discovery and visits to merchant stores. Neither establishes an isolated revenue result for a small merchant in the material reviewed. Chipotle's hiring-time result belongs in the staffing column; Starbucks' companywide performance cannot be credited to Green Dot Assist from the evidence here. For a small business, the opportunity is a shorter, more accurate route to work you already sell. An assistant can help a customer find the right item and reach your checkout, understand a service and request a quote, or repeat a choice without starting from an empty page. Better answers outside staffed hours may reduce missed inquiries. Test those hypotheses against your own results. Track the sequence: useful discovery, visit to the business page, checkout or request started, completed order or staff-confirmed booking, then cancellation or refund if one occurs. A tool call is not revenue. A cart is not revenue. For a quote-based business, even a submitted request may never become a job. Compare channel costs using your own agreements. Your checkout may give you more control over the customer relationship, but it still has payment, fulfillment, support and acquisition costs. A marketplace may supply discovery or delivery you would otherwise need to provide. We have not verified merchant-specific fee schedules, so there is no fee comparison presented as a universal saving. Set a baseline before changing the flow. Compare equivalent periods, record other changes such as promotions and opening hours, and acknowledge missing attribution. A link can carry a general channel label when appropriate; it should not carry a customer identifier or chat transcript. Ask whether the work produces enough completed business or useful staff time to justify maintaining it. ## Risks the owner should recognize **A familiar brand name is not proof of an official connection.** Check the maker field, the connector address and links from the company's own documentation. A domain owned by the brand is useful corroboration; a hosting-provider domain needs additional evidence. Our DoorDash directory observation, for example, did not by itself establish the developer's identity. Use the company's announcement and documentation as the stronger source. **Connecting can share context.** The [DoorDash ChatGPT listing](https://chatgpt.com/plugins/plugin_asdk_app_6943126354348191867f3efffccf94f1) displayed ChatGPT's notice that relevant chats and memories may be shared with a connected app, observed at 18:54:48 UTC on November 7, 2026. Do not interpret "we only ask for an item ID" as a platform-wide guarantee about data sharing. Explain what your own service receives and what the assistant provider says it may send. **A stale answer can create a real customer problem.** Prices, stock and holiday hours need owners and refresh rules. An allergen answer must preserve uncertainty. Make the checkout revalidate the choice, and provide a human route when the facts cannot support an answer. **Payment permission is platform-specific.** OpenAI's [current checkout documentation](https://developers.openai.com/plugins/build/monetization) recommends external checkout for most plugin developers, describes saved-merchant-payment and limited payment-sheet options, and says current commerce approval is limited to physical goods. A generic booking-link design does not establish that every service category qualifies for that platform. Check eligibility before building an integration around it. **Tool descriptions need enforcement.** Read-only claims need matching code and permissions. Retrieved text cannot authorize changing accounts, exporting records or sending messages. Keep staff records separate, and make uncertain or consequential actions visible to the responsible person. ## Two kits to hand to your web person - [Agentic commerce readiness kit (.md)](/downloads/agentic-commerce-readiness-kit.md): business-type checklists, brand patterns, a data worksheet, sign-in decisions, staff helpers, a 30-day plan and a revenue worksheet. - [Order handoff and catalog feed kit (.md)](/downloads/order-handoff-and-catalog-feed-kit.md): original JSON examples for menus, products, locations, structured data and read-only tools, plus choice-only order and booking links and a test list. **Structured data** labels the facts on a page so software can understand their meaning. The feed kit uses Schema.org's vocabulary for menus, products, businesses and offers, and **JSON-LD**, a text format for those labels. Its examples are fictional and must be replaced with your facts. Parsing them verifies syntax and internal consistency. A working store integration and directory eligibility need separate checks. For the wider process of turning a useful small offer into a working business, my book [The $97 Launch](/books/) fits this stage. Start the kit with one customer task and its existing finish line before spending money on another system. ## Related reading - [MCP directories and listing steps](/blog/mcp-directories-listing-guide-2026/) - [Build a connector for businesses, nonprofits, events and WordPress](/blog/build-mcp-connector-nonprofits-events-wordpress/) - [Claude Code and Codex projects: what to borrow](/blog/claude-code-codex-projects-2026-what-to-borrow/) - [Devpost AI projects and real revenue](/blog/devpost-ai-projects-real-revenue-2026/) - [What getting an MCP connector listed takes](/blog/mcp-connector-listing-small-business/) - [Forms the assistant fills and the person sends](/blog/lodging-sites-ai-agents-booking/) - [Plan the website, facts and customer action together](/blog/site-launch-design-search-ai/) - [Grocery brands and agentic checkout: what a small brand can adopt](/blog/grocery-brands-acp-ucp-agentic-checkout-small-business/) ## Fact-check notes and sources Checked November 7, 2026. Announcement dates identify the source's publication; current access can vary by account. Directory observations are separate from company documentation. We did not call another company's MCP tools or complete an order. No third-party code was copied or proposed for reuse; the kits contain original examples. - **Directory observations:** [Starbucks](https://chatgpt.com/plugins/plugin_asdk_app_69cf433334cc8191951af77fbe29f96c), 18:54:26 UTC; [DoorDash](https://chatgpt.com/plugins/plugin_asdk_app_6943126354348191867f3efffccf94f1), 18:54:48; [Instacart](https://claude.ai/directory/instacart), 18:53:10; [Domino's India](https://claude.ai/directory/dominos-india), 18:53:15; [Amazon Selling Partner](https://claude.ai/directory/amazon-selling-partner), 18:53:20; [Uber Eats](https://claude.ai/directory/uber-eats), 18:53:36. Negative search observations describe that browser session only. - **DoorDash:** [grocery app](https://ir.doordash.com/news/news-details/2025/DoorDash-Launches-Grocery-Shopping-App-Within-ChatGPT/default.aspx), read 19:02:59 UTC; [corporate connector](https://about.doordash.com/en-us/news/doordash-ordering-connector), 19:01:54; [developer instructions](https://developer.doordash.com/en-US/docs/mcp/tutorials/get_started/), 19:02:09; [merchant tools](https://about.doordash.com/en-us/news/doordash-unveils-ai-powered-tools-to-enhance-online-menus-and-streamline-merchant-operations), 19:00:37; [DashBuddy](https://about.doordash.com/en-us/news/dashbuddy), 19:01:18; [quarterly shareholder letter](https://ir.doordash.com/news/news-details/2026/DoorDash-Releases-Second-Quarter-2026-Financial-Results/default.aspx), 19:10:15 and 19:10:26. - **Amazon:** [seller listing](https://claude.com/marketplace/connectors/amazon-selling-partner), read 19:00:38 UTC; [AWS MCP](https://aws.amazon.com/about-aws/whats-new/2026/05/aws-mcp-server/), 19:00:37; [AgentCore](https://aws.amazon.com/about-aws/whats-new/2025/10/amazon-bedrock-agentcore-available/), 19:00:55; [Alexa+](https://www.aboutamazon.com/news/devices/new-alexa-generative-artificial-intelligence), 19:01:19; [shopping and historical conversion results](https://www.aboutamazon.com/news/retail/amazon-agentic-ai-gen-ai-shopping), 19:00:02; [Shop Direct](https://www.aboutamazon.com/news/retail/amazon-shop-direct-merchant-workspace), 19:01:54; [Buy for Me](https://www.aboutamazon.com/news/retail/amazon-shopping-app-buy-for-me-brands), 19:00:56; [outside-agent statement](https://www.aboutamazon.com/news/company-news/amazon-perplexity-comet-statement), 19:01:33. - **Starbucks:** [pilot](https://about.starbucks.com/press/2025/meet-green-dot-assist-starbucks-generative-ai-powered-coffeehouse-companion/), read 19:02:40 UTC; [AI update](https://about.starbucks.com/press/2026/supporting-the-moments-that-matter-with-artificial-intelligence/), 19:03:12; [earnings transcript, page 5](https://s203.q4cdn.com/326826266/files/doc_financials/2026/q1/SBUX-1Q26-CORRECTED-TRANSCRIPT.pdf), 19:00:37; [ChatGPT beta announcement](https://about.starbucks.com/stories/2026/meet-the-beta-starbucks-app-in-chatgpt-a-new-way-to-discover-your-next-favorite-drink/), 19:10:10. - **Chipotle:** [hiring launch](https://newsroom.chipotle.com/2024-10-22-CHIPOTLE-INTRODUCES-NEW-AI-HIRING-PLATFORM-TO-SUPPORT-ITS-ACCELERATED-GROWTH), read 19:00:02 UTC; [hiring follow-up](https://newsroom.chipotle.com/2025-02-19-CHIPOTLE-AIMS-TO-HIRE-FUTURE-BILLION-DOLLAR-BUSINESS-LEADERS-IN-NEW-BURRITO-SEASON-HIRING-PUSH), 19:00:37; [kitchen tests](https://newsroom.chipotle.com/2024-09-16-CHIPOTLE-DEBUTS-AUTOCADO-AND-THE-AUGMENTED-MAKELINE-BY-HYPHEN-IN-RESTAURANTS), 19:01:19. - **Other commerce examples:** [Instacart help](https://www.instacart.com/help/section/2893565984/349218736), read 19:03:31 UTC; [OpenAI partnership](https://openai.com/index/instacart-partnership/), 19:03:18; [Walmart](https://corporate.walmart.com/news/2026/10/sparky-shopping-assistant-expansion), [Shopify](https://shopify.dev/docs/apps/build/storefront-mcp) and [Square](https://developer.squareup.com/docs/mcp), 19:00:46; [Toast](https://pos.toasttab.com/news/toast-expands-toast-iq-smart-ai-assistant), 19:03:37. - **Protocol and platform boundaries:** [MCP architecture](https://modelcontextprotocol.io/docs/learn/architecture), read 19:07:05 UTC; [Anthropic everyday connectors](https://claude.com/resources/articles/connectors-for-everyday-life), 19:00:47; [OpenAI checkout documentation](https://developers.openai.com/plugins/build/monetization), 19:01:26. The feed kit supplies its Schema.org and tool-definition sources and parser checks. - **Our implementations:** the public endpoints linked in the small-business section were checked at 18:23:47 UTC, with selected information tools used at 18:24. The [connection guide](/connect/) and [MCP privacy section](/privacy/#mcp) explain our own website-tools connector separately. No conversion, booking revenue or lead-growth result was measured for our examples. Brand and platform research covered 32 pages: 24 read directly and eight completed through a normal signed-in browser, with none left unread. The feed kit adds eight directly read documentation pages. Supplied directory observations and our own connector checks are identified separately above. Some directory and company pages required that browser route. No access restrictions were circumvented. Company-reported figures are labeled; undisclosed revenue, merchant fee schedules and untested account behavior remain unverified. An independent accuracy review re-read the directly readable company and documentation pages between 19:16:01 and 19:16:28 UTC the same day and corrected the wording of five claims. *This post is informational, not legal advice. Mentions of third parties are nominative fair use. No affiliation or endorsement is implied.* --- Canonical HTML: https://jwatte.com/blog/big-brands-mcp-agentic-commerce-small-business/ RSS: https://jwatte.com/feed.xml JSON Feed: https://jwatte.com/feed.json Hero image: https://jwatte.com/images/big-brands-mcp-agentic-commerce-small-business.webp