How to Use OpenCommerce (Agentcard's Buy Tool): The Beginner's Guide

CodexFuse Team · 2026-11-02T23:45:16.651Z · 13 min read min read

What Is OpenCommerce, Exactly?

OpenCommerce is Agentcard's buy-tool SDK for AI agents: one endpoint that gives an agent catalog search, cart, checkout, and order tracking across 5,026 online stores. Its tagline is "OpenRouter for commerce: one endpoint, every online store" — the analogy being that, the way OpenRouter gives developers one API to reach many AI models, OpenCommerce gives an agent one tool to buy from thousands of merchants. (The launch page rounds the store count to the "OpenCommerce 5000 stores" figure.) That tagline is about the idea, not a corporate link: it has nothing to do with OpenRouter the AI model API company, which is a separate business entirely. If you searched "how to use opencommerce" and landed in a tangle of Mailchimp Open Commerce install guides, osCommerce tutorials, and OpenCart setup pages, you are looking at name collisions. As of November 3, 2026, no independent opencommerce tutorial exists beyond the vendor's own quickstart — this guide is about Agentcard's product, opencommerce.sh, which launched on October 29, 2026 — a different thing from the old open-source Mailchimp Open Commerce project on GitHub, from osCommerce, from the "OpenAI Commerce" Magento feed module, and from opencommerce.ir, a separate organization selling open-source agentic-commerce tooling.

Why Can't My Agent Just Buy Things on Its Own?

Your agent cannot complete a checkout without money. That is the whole problem OpenCommerce exists to solve. An LLM agent can browse a store, compare products, fill a cart, and click "checkout" — and then it hits a wall: it has no card, no funds, no way to move actual money. Developers end up building browser-automation hacks, screenshotting checkout pages, or handing the human a link to finish paying by hand. OpenCommerce hands the agent a funded virtual card and a single buy tool, so the purchase flow stays inside the agent's tool-call loop. One call describes what to buy; the endpoint searches catalogs, builds a cart, pays with an Agentcard virtual card, and reports tracking — and a webhook later tells you the charge actually settled, rather than trusting the agent's claim that it paid.

What Does the Buy Tool Actually Do?

The OpenCommerce buy tool is a single tool you hand to your agent, covering the four commerce verbs: search, cart, checkout, and track. This OpenCommerce Agentcard guide follows the official surface: import { AgentcardClient, buyTool } from "@agent-cards/commerce-sdk", then const buy = buyTool({ agentcard, externalUserId, sendLink }), and that buy tool drops into the tool runners of the Vercel AI SDK, Anthropic, OpenAI, or OpenRouter frameworks. (The OpenRouter mention here is about running the agent itself through that model API; OpenCommerce the product and OpenRouter the company are unrelated, as said above.) Once the tool is in the agent's loop, a request like "buy me two AA battery packs from the cheapest covered store" becomes: catalog search across covered merchants, a cart assembled at the chosen store, checkout executed with a single-use virtual card, and order tracking surfaced back through the same endpoint. Agentcard says it searches catalogs and completes purchases in under 20 seconds — a vendor claim — and that purchase-flow tests run continuously at each monitored store with the results published publicly at monitoring.opencommerce.sh, including 45-day test histories, so you can check whether a specific store's flow is healthy before your agent ever touches it.

How Do I Run My First Catalog Search?

Start with a read-only search so nothing can spend money while you learn the shape of the tool. First, create an Agentcard account and get your client credentials, then install the SDK and wire the buy tool into your agent with the buyTool call shown above, passing your Agentcard client, your user's external ID, and a link callback. Next, send the agent a search-only instruction — "search covered catalogs for a USB-C hub under $30 and report the three cheapest options with store names and prices" — and watch the tool calls: the buy tool hits the catalog search endpoint, returns structured product results, and never touches checkout because you never asked it to complete a purchase. Check the returned store names against monitoring.opencommerce.sh to confirm each one's checkout flow is currently passing, since a store can be covered but temporarily failing its public flow test. Finally, before you allow any real money, set the card limits and single-use behavior described in the production section below; an REST API tester is a handy way to probe the underlying endpoints directly and see the raw request and response shapes while you build.

How Do I Run My First Test Checkout?

Run your first purchase on the test storefront, never on a live store. Agentcard's issuing docs name the test merchant as shop.agentcard.sh, which runs in Stripe test mode: test cards are issued instantly, the reveal/close/webhook flows are identical to live, but no real merchant would accept them. The end-to-end arc:

1. Create the client and the buy tool with your Agentcard credentials and buyTool({ agentcard, externalUserId, sendLink }), using a test external user ID so test purchases stay separable.
2. Give the agent a bounded buy order: "Buy the cheapest pack of AAA batteries on shop.agentcard.sh, total under $10." Bounded price and quantity keep the first run cheap and predictable.
3. Watch the four steps in the tool trace: catalog search on the test storefront, cart assembly, checkout with a freshly issued single-use virtual card, and the order confirmation with tracking.
4. Confirm the single-use card closed itself in your Agentcard dashboard after the approved charge — the docs say issued cards close with closed_reason: used after the first approved charge, so a leaked credential cannot be reused.
5. Listen for transaction.authorized on your webhook endpoint — the docs phrase it as "the agent saying it paid is a claim. transaction.authorized is a fact" — and only then treat the purchase as complete in your own records.

What Happens in Production: Store Tests, Cards, and Limits?

Production is where the guardrails matter more than the tool call. First, treat the public monitoring page as your availability signal: opencommerce.sh publishes per-store purchase-flow test results at monitoring.opencommerce.sh with 45-day histories, so check your store's flow status in code or in the dashboard before sending the agent to buy — a covered store with a failing flow test is a store you should avoid until it recovers. Second, lean on the card model: single-use cards close after the first approved charge, which is the default you want for autonomous purchases, while multi-use cards exist for recurring flows and can be managed with pause_card, resume_card, update_card_limit, and close_card operations. Third, pre-set spending limits rather than trusting the agent's restraint: the issuing docs describe per-card limits and settled spend earning 1 token per cent, so cap each card at a sensible ceiling for the purchase at hand. Agents that run this whole loop typically live in a git repository you deploy somewhere — our GitHub beginner guide covers the repo side, our VPS hosting guide covers the server it will run on, and our Docker beginner guide covers containerizing it so the agent, its keys, and its card credentials move as one unit.

How Do I Know the Payment Actually Happened?

Do not trust the agent's word — wait for the transaction.authorized webhook. The OpenCommerce webhook is the single source of truth for the OpenCommerce agent checkout flow: agents narrate their own success, and an agent that believes it completed a checkout may have hit an error it misread, a card decline it didn't notice, or a session that timed out mid-charge. Agentcard's webhook fires "when money moves": on a charge approval, your endpoint receives transaction.authorized with the authoritative details of what was actually charged, and that event — not the agent's "I paid" message — is what your order records should be built on. The practical pattern is: expose a webhook receiver, match the incoming event to your purchase record by the card or transaction ID, and only mark the order "paid" when the webhook lands; if it doesn't land within your timeout window, query the transaction state through the API before retrying the purchase. This is also why the first-checkout arc above ends at the webhook step: the whole point of OpenCommerce is that the agent does the shopping while your systems keep the accounting honest.

OpenCommerce vs Crossmint: What's the Difference?

The closest comparison member in this space is Crossmint's Agent Commerce Toolkit, launched around October 29, 2026 — a single-integration agent payments toolkit that lets AI agents pay with any card and complete checkout at any online store, combining a card vault with Visa Intelligent Commerce / Mastercard Agent Pay credentials and merchant checkout in one API. Per the independent roundup on aissist.io, Crossmint picks a checkout path per purchase (embedded Shop Pay, Universal Commerce Protocol, or browser automation fallback) and also acts as merchant of record while handling chargebacks. OpenCommerce takes the opposite architectural bet: one buy tool backed by Agentcard's own checkout execution, with per-store purchase-flow tests published publicly so coverage is verifiable before you buy. The table below is an original comparison built from both vendors' public pages as of November 3, 2026:

What Does OpenCommerce Cost?

OpenCommerce pricing has two tiers as published on opencommerce.sh: Standard at 1% of each completed order, which is the tool-only tier with no direct API access, and Enterprise with custom fees plus API access. Billing runs through your Agentcard account via a prefund balance or a pay-as-you-go card, charged at 1% of each live purchase. To make the 1% concrete, here is a computed example set at the Standard rate — these are arithmetic, not quoted prices:

The practical habit: set each single-use card's limit a few percent above the expected total so the fee and any tax don't push a legitimate purchase over the cap and trigger a decline mid-checkout. No pricing was listed on the Product Hunt launch page, and the vendor's pages are the only current source for these figures — verify them before budgeting, since the product is days old and terms can move.

7 Mistakes Beginners Make with OpenCommerce

1. Trusting the agent's "I paid" message instead of the transaction.authorized webhook — build records on the webhook, never on chat output.
2. Skipping the test storefront — every first purchase belongs on shop.agentcard.sh in Stripe test mode, not on a live store with a real card.
3. Mixing up the names — "OpenCommerce by Agentcard" is the buy-tool SDK; the old Mailchimp Open Commerce, osCommerce, and OpenCart are unrelated e-commerce software.
4. Sending the agent to a store with a failing flow test — check monitoring.opencommerce.sh for your store before the purchase, not after the failure.
5. Leaving cards uncapped — pre-set limits and use single-use cards by default so one confused agent can't spend the whole budget.
6. Pasting real credentials into chat or code snippets — keep Agentcard keys in environment variables or a secrets manager, and use test credentials everywhere in guides.
7. Forgetting the 1% fee in budget math — a card capped at exactly the order total will decline once the fee lands; add margin.

Sources

• OpenCommerce on Product Hunt (tagline "OpenRouter for commerce: one endpoint, every online store", maker Pilar, published October 29, 2026, page updated November 2, 2026; 5,000+ stores and under-20-second purchases are vendor claims) — https://www.producthunt.com/products/opencommerce
• opencommerce.sh official site (5,026-store figure and under-20-second purchases are vendor claims; SDK surface @agent-cards/commerce-sdk with buyTool({ agentcard, externalUserId, sendLink }); Vercel AI SDK / Anthropic / OpenAI / OpenRouter tool runners; monitoring.opencommerce.sh 45-day histories; Standard 1% per completed order, Enterprise custom; billing via prefund balance or pay-as-you-go card) — https://opencommerce.sh/
• Agentcard issuing docs: Using the Cards (page updated October 23, 2026 — test storefront shop.agentcard.sh in Stripe test mode; test cards issued instantly, not accepted by real merchants; buy tool on the same MCP connection; transaction.authorized webhook "when money moves"; single-use cards close with closed_reason: used; multi-use cards via pause_card/resume_card/update_card_limit/close_card; 1 token per cent of settled spend) — https://docs.agentcard.com/issuing/using-the-cards
• Crossmint Agent Commerce Toolkit coverage via the independent roundup on aissist.io (merchant of record, chargeback handling, embedded Shop Pay / Universal Commerce Protocol / browser-automation fallback) — https://aissist.io
• my2cents.ai roundup, October 31, 2026 (independent editorial coverage of the launch) — https://my2cents.ai
• fintechize Substack, November 2, 2026 (newsletter mention) — https://fintechize.substack.com
• M11 Labs OpenMarket research preview (openmarket.m11.ai — multi-agent marketplace demo surface where the buyer agent cannot purchase; comparison member, never a rival) — https://openmarket.m11.ai

Frequently Asked Questions

Is OpenCommerce the same as the old Mailchimp Open Commerce?

No. Agentcard's OpenCommerce (opencommerce.sh) is a buy-tool SDK for AI agents launched October 29, 2026. Mailchimp Open Commerce (github.com/mailchimp/opencommerce) is an older open-source e-commerce platform, and osCommerce and OpenCart are classic e-commerce software packages — all unrelated to this product.

How much does OpenCommerce cost?

Standard is 1% of each completed order (tool only, no API access); Enterprise has custom fees plus API access. Billing runs via a prefund balance or pay-as-you-go card. The Product Hunt launch page listed no pricing, so confirm the current terms on opencommerce.sh before budgeting.

Does OpenCommerce work with any AI agent framework?

Per the official quickstart, the buy tool drops into the tool runners of the Vercel AI SDK, Anthropic, OpenAI, and OpenRouter frameworks. The underlying SDK is @agent-cards/commerce-sdk.

Can my agent really buy from any of the 5,026 stores?

The 5,026-store figure is a vendor claim, and coverage is not the same as reliability: opencommerce.sh publishes public per-store purchase-flow test results at monitoring.opencommerce.sh with 45-day histories, so verify your specific store's flow is passing before sending the agent to buy there.

Is OpenCommerce related to OpenRouter?

Only by analogy. "OpenRouter for commerce" is Agentcard's tagline, meaning one endpoint for many stores the way OpenRouter is one API for many AI models. The two companies are separate and unrelated.