YuMoney Launches MCP Server for Payments and Explains How to Write APIs for Neural Networks
YuMoney has launched an MCP server, which lets a neural network call the payment service's functions directly through the Model Context Protocol. The team found that without a line saying "do not make up the amount, ask the user," the model would sometimes create payments for invented amounts. That is why parameter descriptions and error texts are written as instructions for the model. The takeaway: in MCP, the description field works as part of the prompt and determines whether the integration works at all.
AI-processed from Habr AI; edited by Hamidun News
YuMoney has launched an MCP server for payments and invoices. MCP (Model Context Protocol) is a protocol through which a large language model (LLM, a ChatGPT-like neural network) calls functions of an external service on its own: create a payment, send a receipt, check the status. Previously, such functions were exposed through a regular REST API, a set of requests that code written by a human would call. Backend developer Roma at YuMoney has spent nearly a decade developing the public payment API, and he explained that an interface for a neural network has to be designed differently. The main example: without a single line in the parameter description, the model sometimes created a payment for an invented amount.
The Parameter Description Is a Prompt
In a classic API, documentation is written for people. If a field is poorly described, the developer will check a forum or write to support. In MCP, the tool description effectively becomes part of the system prompt, that is, an instruction the model follows literally. That is why naming and clarity of the description determine whether the integration works.
In YuMoney's regular documentation, the payment amount is described through nuances: fees from YuKassa partners, currency codes, matching the currency of the sub-account or the store. For MCP, the description was shortened and a direct instruction was added: the value must not be invented, and if the amount is not specified, the user must be asked for it. It is obvious to a human that the customer names the amount, but not to a model. Without that line, the model created payments for made-up amounts.
The same logic applies to errors. A response like insufficient_funds is convenient for code, but for a model it is a dead end. In an MCP response, the team adds text: "Insufficient funds to complete the operation. Inform the user. Do not try to make the payment again." A clear error lets the model either fix the request and resend it, or stop and honestly explain the situation.
Another trap is overloaded schemas. If a request contains similar fields such as billing_address, shipping_address, and user_address, a human will sort them out and check before launch, but a model may put the data in the wrong place. The author advises avoiding fields with close meanings, keeping only the most necessary scenarios in the server, or describing in detail which field to fill in which case.
Security and Repeated Calls
An MCP server does not know who is calling it: a powerful cloud model or a small local one with 7 billion parameters (7B). Weak models confuse the order of calls and repeat requests at the slightest doubt. Hence the rule: all methods must be idempotent, meaning a repeated call must not charge money a second time. YuMoney goes further for now and does not allow financially significant operations through MCP at all without human involvement.
The author advises separating MCP authorization from regular API authorization. Otherwise, a model trying to complete a task may go to the classic API with its own key. In addition, incoming data must be checked more strictly than in a regular API: the model may generate an invalid or even malicious request.
The myth of "freedom from versions" is also debunked. In MCP, the list of tools is updated on the fly through tools/list, and the model rereads the schema each time. But a function in use still cannot be deleted: clients' business processes will break, even if the client is a neural network.
Tokens, Integration, and Latency
In a regular API, the size of a JSON schema matters only for traffic. In MCP, every function description and every field enters the model's context window, spending the user's tokens and money. Concise schemas and responses directly reduce the cost.
Integration changes too. In REST, you expose atomic methods, and the client writes code: create a payment, send a receipt, confirm the operation. In MCP there is no rigid sequence; the model chooses the order of calls based on the request and the dialogue context. Typical scenarios can be fixed by the user through Rules and sets of instructions (Skills) in their AI agent, without writing a single line of code. According to the author, the cost of integration through an LLM tends toward zero.
Latency is more complicated. The user's wait consists of the model's "thinking," network delays, and server-side processing of the request, and the cycle can repeat with clarifications and errors. The server developer mainly influences the last item and indirectly the number of iterations: the clearer the descriptions, the fewer unnecessary attempts.
What Follows
The article is a practitioner's view, not a study: it contains no measurable metrics, and "the cost of integration tends toward zero" remains the author's forecast. Still, the set of rules is workable for any team preparing an MCP server. Write descriptions as instructions, make error messages conversational, trim schemas, make methods idempotent, separate authorization, and keep a human in the loop for money operations. The author expects large companies to adopt MCP more actively in B2C, primarily where a model error will not lead to disaster. YuMoney plans to develop the functionality further and is planning a public online hackathon.
- Nov 21, 2026OpenAI Names 4 Reasons GPT-6 Astra Hides Its Reasoning from Monitors
- Nov 21, 2026Australia's Central Bank: AI Hits Young Workers Aged 15–24 First in IT and Customer Support
- Nov 21, 2026Claude Code Taught to Hunt Vulnerabilities: 83 Skills Built from 681 HackerOne Reports
- Nov 21, 2026DeepSeek Weights Show Traces of NVIDIA Blackwell GPUs: 445 of 469 Rows Match on B300
Key Questions Answered
What are the key details on The Parameter Description Is a Prompt?
In a classic API, documentation is written for people. If a field is poorly described, the developer will check a forum or write to support. In MCP, the tool description effectively becomes part of the system prompt, that is, an instruction the model follows literally. That is why naming and clarity of the description determine whether the integration works.
What are the key details on Security and Repeated Calls?
An MCP server does not know who is calling it: a powerful cloud model or a small local one with 7 billion parameters (7B). Weak models confuse the order of calls and repeat requests at the slightest doubt. Hence the rule: all methods must be idempotent, meaning a repeated call must not charge money a second time. YuMoney goes further for now and does not allow financially significant operations through MCP at all without human involvement.
What are the key details on Tokens, Integration, and Latency?
In a regular API, the size of a JSON schema matters only for traffic. In MCP, every function description and every field enters the model's context window, spending the user's tokens and money. Concise schemas and responses directly reduce the cost. Integration changes too. In REST, you expose atomic methods, and the client writes code: create a payment, send a receipt, confirm the operation. In MCP there is no rigid sequence; the model chooses the order of calls based on the request and the dialogue context.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.