232 Russian services have APIs for AI agents, but only five allow payments

⚡ The gist in 5 seconds

• Out of 424 Russian e-commerce and fintech services, 232 have integrations for AI agents — 312 in total (API, MCP, A2A).
• Of 50 integrations for personal tasks, only 17 have confirmed official standalone access, and just 5 can actually process a payment.
• Even where an official MCP exists, you can't buy anything: VkusVill hands over a cart link, while checkout and payment remain with the human.

🔍 What happened

A Habr author under the nickname dochkinavika examined the public documentation of 424 Russian companies and products in e-commerce and fintech and calculated which of them a personal AI agent could actually connect to. She deliberately set aside the connection method: what matters more is which operations a service exposes to the agent and on what terms. Integrations were found at 232 services — 312 in total, with some companies offering several. The full breakdown and a map of services are in the article on Habr.

A breakdown by purpose: 233 integrations are aimed primarily at businesses, 50 at personal tasks. Of those 50, official standalone access is confirmed for 17 — that is, 14 services. Actions with a physical result — placing an order, arranging delivery, commanding a device — are declared in 10 personal integrations, and half of those were built by third-party developers rather than the services themselves.

The mechanics are visible in three examples. VkusVill has an official MCP: the agent finds products and calls the vkusvill_cart_link_create method, which returns a cart link. Authorized users get access to order history and personalized offers, but the published set contains no checkout or payment methods. Lenta has an open catalog API — stores, products, prices, availability, no authorization required — but checkout and payment could not be confirmed in the published set.

Delovye Linii let an agent create a shipping request and request changes, but cancellation requires verification and can be rejected, and a payment link doesn't mean a completed payment. In a sample of 59 integrations with personal scenarios, 53 support search, 8 support creating or confirming an order, and 5 support a payment operation. Of the five payment capabilities, four are third-party MCPs and one is the official YuMoney API. Another 29 integrations the author classified as data and development. The Ozon Seller API lets you manage products, prices, and shipments — but that's a seller API, not a buyer's.

🧠 Our take

The figure "232 out of 424" looks like a boom, but it needs to be read soberly. Nearly half of the integrations are a merchant's storefront: export products, update prices, view shipments. Our assessment: businesses connect agents where a contract and invoice-based payment already exist, because that's easier to negotiate. The consumer scenario requires something else — a person's consent to charges, refund disputes, liability for errors. Until that exists, MCP remains a pretty wrapper around a catalog.

What this changes for your wallet and choices — very little, for now. An agent genuinely saves time at the "find and compare" stage: 53 of 59 integrations support search, and that's already useful. But the final step — the very one this was all supposed to be about — is still done manually by a human in the app. It looks like a transitional period: the infrastructure has arrived, but the payment fork hasn't.

Fintech and platforms are winning: integrations were found at 48 of 56 fintech services and 46 of 65 in store and customer management, versus 9 of 19 in entertainment and 4 of 9 in home services. We believe the first full-fledged agentic payment will come not from retailers but from those who already have a charge API and a license. Small shops lose out: they have neither the team nor the resources to write an MCP, and they simply won't appear in agent results.

⚠️ What we don't know

The main weakness is the method: public documentation was counted, not live integrations. Whether a particular third-party MCP actually works, how fast it responds, and what it does on error doesn't follow from the article. Half of the personal integrations with physical results were built by third-party developers — there's no data on their reliability or data handling. The composition of the 424 services themselves is also unclear: how they were selected and who was left out. It's worth checking limits, access terms, and the existence of a partnership agreement before building a product on this.

Prediction: within a quarter, at least one major Russian grocery retailer will announce checkout via MCP or an agentic API — but with payment confirmation by a human.

Disclaimer: a single source — an independent study by one author based on public documentation; live integrations were not tested.

Source: Habr