Walmart opens AI checkout, with permission controls attached

Walmart is giving approved AI agents a supported way to search its inventory and buy products through a new MCP server called Agent Gateway. It’s also introducing account-level controls that let shoppers approve individual purchases or authorize spending within a limit.

Those controls deserve close attention. Connecting a model to a shopping tool is relatively straightforward. Deciding whose money it can spend, under what authority, and with what recourse is harder.

Walmart announced the technology on October 21, 2026, with support planned for Walmart and Sam’s Club. Agent Profiles, the customer-facing authorization mechanism, are scheduled to roll out in November. According to TechCrunch’s report, published October 21 and updated October 22, shoppers won’t need Walmart+ to use the agent support. A Walmart account is sufficient.

The gateway gives developers a sanctioned transaction path as retailers become more restrictive about agent access. Amazon has blocked AI agents from making purchases, while some travel and booking sites have shut agents out altogether.

Walmart’s access is conditional. Agents need approval, must follow specific terms, and can be blocked for violating them.

Connecting agents to Walmart’s transaction system

Agent Gateway uses the Model Context Protocol, originally created by Anthropic, to connect external AI agents to Walmart’s shopping capabilities. The announced functions include inventory search and purchases on a user’s behalf.

A supported tool interface gives engineers a more predictable integration target than a retailer’s consumer website. Browser agents have to interpret pages, locate controls, cope with layout changes, and determine whether checkout succeeded. A purpose-built interface can reduce that ambiguity.

The report doesn’t include Walmart’s tool schemas or transaction flow, though, so there’s little basis yet for judging the implementation’s completeness or reliability.

MCP doesn’t settle authorization by itself. It provides a standard way to connect agents with tools and data. Walmart still has to decide which agents may connect, what customer information they receive, and which purchase requests they may execute.

A model can produce a plausible shopping request without having permission to spend money. The transaction system needs to enforce the shopper’s authority independently of whatever the agent says. Walmart plans to handle those decisions through Agent Profiles and operating terms.

Enforcing restrictions at the system that holds the account and processes the purchase is a sensible choice. Relying on an agent to obey a budget offers a weaker safeguard.

Walmart already has a partnership with Meta’s Muse and integrations with apps including ChatGPT and Gemini. The gateway adds a common protocol-based route, though the report doesn’t establish that every existing integration uses it.

Delegated spending becomes an account feature

Agent Profiles will give each authorized agent a dedicated profile linked to the shopper’s Walmart account. Customers create and link that profile when first shopping through the gateway, then choose the controls:

• Require approval for every transaction.
• Authorize transactions within a customer-set spending limit.
• View agent purchases separately from direct purchases.
• Delete an Agent Profile at any time.

Separate attribution should make mistakes easier to investigate. If several assistants shop for one household, a combined order history can obscure which one placed an unwanted order. An agent-specific record gives the customer somewhere to start.

Walmart compares the arrangement to adding an authorized user to a credit card, with identifiable transactions and capped spending. That’s a useful explanation of delegation, though it doesn’t establish identical financial protections.

Consider an assistant asked to replenish coffee, detergent, and pet food. Per-transaction approval lets the shopper check the proposed purchase before money changes hands. A spending allowance removes that interruption but leaves more decisions to the agent.

An order can stay under budget while selecting the wrong package size or an unwanted substitute. Walmart says agents must stay within the shopper’s intent, but the report doesn’t explain how that requirement becomes enforceable purchase policy.

Teams building shopping assistants still need clear product selection and understandable approval requests, even when a supported API handles checkout.

Personalization without handing over purchase history

Walmart says the gateway will help shoppers buy their usual brands and items without sharing personal information or purchase history directly with the agent.

That suggests a separation between retailer-held context and agent-visible information. Walmart can use information it already holds to support a shopping task while limiting what crosses into another company’s system.

The mechanism hasn’t been disclosed. The report doesn’t say whether the agent receives suggested products, filtered results, or some other representation. Those details will determine how much an external assistant can infer.

Walmart’s terms also say agents can’t access customer payment information and may access only the information needed for the task. Keeping payment credentials outside the agent’s reach limits the sensitive data handled by the assistant and its surrounding infrastructure.

Other exposure remains. Product requests and selected items can reveal preferences, so developers should examine gateway responses and their own logging practices before describing an integration as private. Walmart’s claim concerns avoiding direct disclosure of personal information and purchase history. It doesn’t rule out inferences from shopping activity.

Risks and limitations

Several safeguards are announced rather than demonstrated. Agent Profiles are due in November, and the report provides no implementation documentation, performance measurements, or operating history for the gateway.

Engineers can’t yet assess production readiness. The report doesn’t specify rate limits, latency, availability commitments, or behavior when inventory changes between search and purchase. It also doesn’t explain what happens when a purchase request is retried after a timeout. Uncertainty about whether an order succeeded is a serious problem for a transaction integration.

The spending limit needs more definition, too. The source doesn’t establish whether it applies per transaction, over a time window, or through another accounting model. Those semantics matter when an agent makes multiple purchases.

Access remains conditional on Walmart’s approval and terms. MCP support doesn’t guarantee that a developer can connect any assistant. Teams need to understand the approval process and access conditions before committing to an integration.

Walmart says its returns process applies when an agent doesn’t act as intended. That gives shoppers a remedy, but the report doesn’t describe separate compensation for agent errors or exceptions to ordinary return rules. Returning an unwanted product still takes effort.

“Shopper intent” is also a demanding standard. A spending cap is measurable; whether a substitution satisfies a user’s request can be harder to determine. The announcement doesn’t explain how Walmart will resolve disputes between transaction policy and natural-language instructions.

What developers should check next

Walmart is putting delegated shopping into the customer account, with explicit permissions, traceable purchases, and a returns process. Those are useful foundations for agents that can spend real money.

Technical teams should examine the transaction contract before building around it: authorization scope, approval state, retry behavior, spending-limit semantics, and the data returned to the assistant. Each affects whether an agent can complete purchases safely.

Retailer cooperation makes that work easier. A supported gateway gives developers a clearer integration target and gives Walmart control over access and enforcement. Amazon’s blocking approach leaves agent builders with fewer sanctioned options.

The November rollout should provide a better basis for assessing whether shoppers can delegate routine purchases while keeping track of what they’ve authorized. The quality of the permission controls, and how reliably Walmart enforces them, will matter as much as checkout itself.