Head to head · Placing and tracking orders · October 2026 research run

CartAI vs opencommerce

opencommerce scores 57.9 (C) on agent readiness against CartAI's 34.9 (F), and leads in every scored category. Both do placing and tracking orders.

Best shopping and order APIs for AI agents buying for a person · All 61 shopping comparisons

Which one, for what

CartAI F

Best for Platforms that show products (rewards apps, publishers, vertical AI apps) and want to complete purchases on US merchants' own sites without merchant integrations.

Also in its favour

  • Runs on your own machine
  • Free to start without a card

Watch for

No public prices. Production fees are basis points on order value, shown only in the dashboard

opencommerce C

Best for A company building a chat agent that buys for US shoppers at many online stores, restaurants and airlines through one tool, with payments approved by the shopper's passkey.

Ahead on

  • Reliability, 25 against 18
  • Schema & documentation, 81 against 59
  • Agent ergonomics, 69 against 49
  • Security & auth, 68 against 43
  • Payments & pricing, 60 against 30
  • Maintenance & community, 75 against 18
  • Transparency & trust, 63 against 45

Also in its favour

  • A hosted endpoint, with nothing to install

Watch for

The tool description is 57 rules and about 14,500 characters, read on every call

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Maintenance & community opencommerce +57

CartAI 18

Closed service with no changelog. The newest dated release found is the MCP server's last commit on 20 July 2026, 123 days ago (10 of 30). Docs pages show updatedAt dates from 17 to 20 November 2026, but no dated change list, so no releases can be counted in the last 90 days (0 of 20). No public changelog. Support is through a contact page, info@cartai.ai and a portal request for more credits that the docs say is reviewed within one business day. Neither was tested (5 of 25). Not in the official MCP registry, not on npm, and no official SDKs (0 of 15). The MCP server pins @modelcontextprotocol/sdk ^1.0.0 and zod ^3.22.0 with a lockfile and no CI (3 of 10). The earliest first-hand date of a running API is a webhook example of 6 December 2025, and the MCP repository dates from 25 May 2026. 10 + 0 + 5 + 0 + 3 = 18.

opencommerce 75

@agent-cards/commerce-sdk 0.18.6 was published on 30 October 2026, per the 30 October check, not re-read on 2 November (30 of 30). 19 versions from 16 to 30 October 2026 (20 of 20). Closed service. Agentcard's changelog is dated almost daily, newest 27 October on 2 November too, and support runs through agent-cards support chat and support@agentcard.sh. Neither was tested (10 of 15). The TypeScript SDK is current, with no SDK in a second language (10 of 15). The SDK has no dependencies. The CLI has 10, among them PostHog and the Claude Agent SDK, needs Node 22, and its repository is private, so CI isn't visible (5 of 10). The product is two weeks old. Total 75.

Payments & pricing opencommerce +30

CartAI 30

No x402, MPP or L402. The card sessions pay merchants for customers' orders, not CartAI (0 of 10). No price list without a login. The terms say production fees are basis points on order value and other usage metrics, shown in the dashboard or agreed separately, and credits have no published price (0 of 30). Every new account gets 5 free credits, about 500 Catalog API calls or at least 10 test checkouts, and the signup asks for name, email and password with no card (30). A person signs up in a browser to get a key, and production needs a Go Live review (0). 0 + 0 + 30 + 0 = 30.

opencommerce 60

No x402, MPP or L402. Purchases are paid with the shopper's card from the Vault, and agentcard.com's FAQ answers Not yet on agent payment protocols (0 of 40). 1 per cent of each completed order is published on opencommerce.sh without a login (20 of 20). Sandbox credentials run the whole flow, stop before payment and are never billed, with no card needed (20 of 20). A person signs in through app.agentcard.sh or an emailed code, and since 27 October 2026 the CLI says sign-ups are by invitation (0 of 20). Total 40. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 40 is 60.

Security & auth opencommerce +25

CartAI 43

One API key per request in a header, revocable and regenerated in the portal, with no scopes or rotation guidance found (20 of 30). No read-only key. A checkout can wait for the customer's confirmation, and production payments run through hosted Visa Intelligent Commerce or Mastercard Agent Pay sessions the customer authorises. The MCP server places orders with no confirmation of its own (10 of 20). Product search and details return merchant page content with no prompt-injection guidance, and the create_payment_session tool description tells the model to present the payment link "without asking the user first" (3 of 15). Each task has a status history with timestamps and credits, a live preview URL and webhooks for every status (10 of 15). No security.txt (404), disclosure policy, bug bounty or certification found. The known limitations page says test-mode card data is not tokenised and could appear in logs (0 of 20). 20 + 10 + 3 + 10 + 0 = 43.

opencommerce 68

OAuth client credentials exchanged for a one-hour platform token, separate sandbox and production clients, and a per-shopper access token as an alternative. Agentcard's platform tokens carry the single scope api, per the agentcard listing, and init writes production credentials to the env file by default (22 of 30). Every payment not covered by an auto-approval rule needs the shopper's passkey, the docs say neither the app nor the model approves for the shopper, and the model cannot change the shopper or address. init runs Agentcard's coding agent with permissions bypassed and unsandboxed commands, which the docs disclose. Two less for that (15 of 20). Answers carry store and product text, links are stripped before the model reads them, and preset rules ignore merchant text an agent sends, but no guidance naming untrusted content or prompt injection was found (10 of 15). Order and purchase events by webhook or stream, with logged deliveries and stored approval records of who, when and where (11 of 15). agentcard.com serves a valid security.txt, the Vault Enterprise plan lists a SOC 2 report and pen-test summary, and the Trust Centre was not read. No bug bounty or disclosure policy found (10 of 20). Total 68.

Schema & documentation opencommerce +22

CartAI 59

A public OpenAPI 3.0 file at docs.cartai.ai/openapi/latest.yml with 13 operations. It declares no security scheme, and the key header is spelt X-Api-Key on some paths and x-api-key on others, which CartAI's own auth.md points out (20 of 25). llms.txt on both hosts and a Markdown twin of every docs page (10). Reference pages say when to use each call, such as a table of when to use webhooks or polling for a task, and the MCP tool descriptions say which call comes next (14 of 20). Required fields and patterns for phone and postal code, but only two enums in the spec, shippingMethod.strategy is a free string, and payment.data is typed as a string in one table while the examples send an object (7 of 15). Request and response examples on each page and a 401 body, but most operations document only 400, and a link to the task lifecycle section points to an anchor that is missing (8 of 15). No changelog and no API versioning beyond 1.0.0 in the spec (0). 20 + 10 + 14 + 7 + 8 + 0 = 59.

opencommerce 81

The buy tool's input is a strict JSON Schema with five required fields and additionalProperties false, and the package ships TypeScript declarations. The purchase endpoints under it are missing from the OpenAPI file at docs.agentcard.com/openapi.json, which holds 18 operations for tokens, connections, identity checks and cards (20 of 25). llms.txt on opencommerce.sh and on docs.agentcard.com, with a Markdown twin for each docs page (10 of 10). The description's 57 rules say when to call the tool, when to set confirm and what each answer means, such as "Show the shopper the cart before you confirm it" (18 of 20). Two booleans and two nullable ids, but the main input ask is free text (9 of 15). Code for Anthropic, Vercel AI SDK, OpenAI and OpenRouter, a sandbox walkthrough script, error tables with what to do next and captured error bodies (14 of 15). npm versions follow 0.x semver, and Agentcard's changelog is dated almost daily, newest 27 October 2026, but it covers every Agentcard product and the SDK has no changelog of its own (10 of 15). Total 81.

Agent ergonomics opencommerce +20

CartAI 49

The MCP server has 8 tools with compact descriptions, but each result appends the full JSON response, and the REST API has no field selection (22 of 25). Product search takes a query and an optional merchant to prioritise, with no limit, page or filter parameters found (6 of 20). A documented 401 body with an error code and message, 400 on every operation, and 403, 404 and 409 on interactive inputs. Error bodies for failed checkouts are not described beyond the FAILED status (10 of 20). No idempotency key on writes, and the docs say a resubmit creates a duplicate. A confirmation step (verifyBeforePlacement, PENDING_CONFIRMATION) and a cancel call exist, and the MCP tools carry no readOnlyHint or destructiveHint (6 of 20). A checkout needs a full customer object with contact, two addresses, shipping strategy and payment. No official SDK, and the MCP server runs from a clone, in Node.js only (5 of 15). 22 + 6 + 10 + 6 + 5 = 49.

opencommerce 69

One tool, inside the band for ten or fewer, but its description is 57 rules and about 14,500 characters, and the docs also recommend Agentcard's appInstructions() system prompt. Answers are capped at eight product and eight store cards and stripped of links for the model. Ten less than the band for the description's size, two back for the capped answers, a departure from the checklist (17 of 25). The buy tool has no page or filter controls, and storeWaitMs sets how long a reply waits. GET /buy/merchants has q, kind, limit up to 100 and a cursor (12 of 20). AgentcardApiError codes each come with what to do next, decline_code and error_code are machine-readable, and input outside the schema is refused with a message the model can act on (18 of 20). No idempotency key on a purchase turn. The client never resends a turn except once after a 401, a cart hash guards confirms, order events carry an order_id to reconcile on, and the MCP buy tool is documented as destructive (14 of 20). The model fills five simple fields, but the SDK is TypeScript only, for Node 18 or later and the browser (8 of 15). Total 69.

Transparency & trust opencommerce +18

CartAI 45

Closed service with published terms naming CartAI, LLC under Texas law, updated 13 May 2026, and an Apache 2.0 MCP server (15 of 30). The privacy policy (13 May 2026) describes card vaulting through a partner and says raw card numbers never reach CartAI systems, but the docs say test-mode card data is not tokenised and could appear in logs. Retention is "as long as necessary" with no periods, and no DPA was found (12 of 30). No API deprecation policy. The terms let CartAI change or discontinue parts of the Services without liability and promise notice only for material changes to the terms and to pricing (0 of 20). The privacy policy names Visa Intelligent Commerce and Mastercard AgentPay and other providers by kind, and the terms name Cloudflare, HUMAN and Fingerprint as bot-mitigation partners. Processing is in the US and other countries, with no subprocessor list (6 of 20). 15 + 12 + 0 + 6 = 33.

opencommerce 63

Closed. The packages' licence allows running them only to use Agentcard's services and forbids benchmarking a competing product. The terms of use (10 July 2026) name Agentcard Corporation and cover the API, MCP server and CLI, but not the Commerce SDK or its 1 per cent fee, and the opencommerce.sh footer names Agentcard, Inc. (12 of 30). Agentcard's own privacy policy (21 October 2026) covers orders, delivery addresses, store accounts it links or creates with their encrypted sign-ins, a one-time card credential held at most 45 minutes, and session recordings of what is typed into dashboard forms. Retention is as long as the account is active, with no periods and no public DPA (14 of 30). No deprecation policy. The SDK keeps deprecated names with notes, such as toAnthropicTool and ackFor, and the changelog marks one breaking change of 10 September 2026 (6 of 20). The privacy policy names Crossmint, Coinbase, Third National, Sumsub, Visa and Stripe and the kinds of other providers, with transfers to the United States. The CLI's PostHog telemetry honours DO_NOT_TRACK and AGENT_CARDS_TELEMETRY, found in its code (10 of 20). Total 42.

Score by category

CategoryWeightCartAIopencommerceEdge
Reliability16%1825opencommerce +7
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%5981opencommerce +22
Agent ergonomics13%4969opencommerce +20
Security & auth14%4368opencommerce +25
Payments & pricing10%3060opencommerce +30
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%1875opencommerce +57
Transparency & trust7%4563opencommerce +18
Negative events≤15-3-3
TotalF 34.9/100C 57.9/100

Facts side by side

FactCartAIopencommerce
KindHTTP APISDK + MCP
VendorCartAIAgentcard
Hosted endpointno (local only)https://api.agentcard.sh
TransportsstdioHTTP
AuthAPI keyOAuth
PricingFreemiumPay per use
x402nono
LicencenoneProprietary. @agent-cards/commerce-sdk and the agent-cards CLI ship a licence of Agentcard Corporation that allows running them only to use Agentcard's services and forbids copying, reverse engineering and benchmarking a competing product. The repository they name is private
Tools exposed81
Read-only variant documentednono
llms.txtyesyes
Last release2026-07-202026-10-30
Terms last updated2026-05-132026-07-10
Privacy policy last updated2026-05-132026-10-21
Customer content may train modelsnot found in the textnot found in the text
Terms restrict automated accessyesyes
Terms restrict benchmarkingyesnot found in the text
Terms or service can change without noticenot found in the textnot found in the text
Arbitration or class-action waiveryesyes
Popularitynone2.7k npm/wk

How the scores have moved

Verdicts

CartAI

CartAI lets a platform buy from US merchants' own sites through one asynchronous API, with a public OpenAPI file, Markdown docs, webhooks and an optional customer confirmation step. There is no public price list, status page, rate limit or idempotency key, and the home page's retry-safe and Hosted Cart claims go further than the docs.

opencommerce

One strictly typed buy tool covers search, cart, checkout and order updates, and every payment not covered by an auto-approval rule waits for the shopper's passkey. The SDK is closed source and 0.x with no public CI or issue tracker, its tool description runs to about 14.5 KB, and the purchase API under it is undocumented.

Before you call either

CartAI

  1. Never resend POST /checkout after a 201 or a timeout. Look up the task with GET /checkout/{taskId} first, because a resubmit creates a duplicate order
  2. Use only the test card 4242 4242 4242 4242 until the account is in production mode, since test mode does not tokenise card data
  3. Set verifyBeforePlacement and show derivedDetails.pricing to the customer at PENDING_CONFIRMATION before the order is placed
  4. Poll a task no more than once every 5 to 10 seconds when webhooks are unavailable
  5. Give selectedVariant labels as the merchant's page shows them, and use US merchants, addresses and phone numbers only

opencommerce

  1. Call buy with the shopper's words in ask, and set confirm true only after the shopper says yes to the cart you showed
  2. After a timeout, unreachable or invalid_answer error, read the conversation before sending the turn again, since the order may have gone through
  3. Show the cart's max_charge beside its total, because the approval holds up to that amount
  4. Use sandbox credentials for tests. init writes production credentials by default, and they place real orders
  5. Check a store with GET /buy/merchants before sending a shopper there, and suggest another store when it is held_back or supported is false

Questions

Which is better for AI agents, CartAI or opencommerce?

opencommerce scores 57.9 (C) on agent readiness against CartAI's 34.9 (F), and leads in every scored category.

Can an agent call CartAI and opencommerce without installing anything?

CartAI runs on your own machine, with no hosted endpoint listed. opencommerce has a hosted endpoint at https://api.agentcard.sh.

Other comparisons with CartAI or opencommerce

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.