From REST to MCP: How Language Models Are Changing API Design

Talk of artificial intelligence soon replacing developers and business processes entirely is everywhere. But the technical details are often lost amid the sensational headlines: exactly how AI agents should interact with existing systems, and why traditional approaches to API design don’t always work here.

Most modern services are built on REST architecture, which was originally designed for human-machine interaction via clear, strictly typed endpoints. This approach works great when the client is pre-written code that knows the structure of requests and responses in advance.

The Problem with REST for LLM Agents

When a language model acts as the client, the situation changes. The agent must not only call an API method but also understand which methods exist, what they are used for, and how to combine them to solve a task formulated in natural language.

Standard REST documentation—in the form of OpenAPI specifications or Swagger schemas—is designed for human developers or strictly deterministic code. It is ill-suited for a model to independently construct the logic of actions based on it, especially if the system consists of dozens or hundreds of services.

What MCP Offers

As an alternative, the MCP (Model Context Protocol) is increasingly being discussed — an approach originally designed with the understanding that the primary “client” might not be a human or a rigidly programmed service, but rather an AI agent.

The key idea behind MCP is to provide the model with context and tools in a format that it can interpret and use independently, without the need to define the entire interaction logic in advance. This changes the very principle of API design:

• instead of strict contracts—flexible descriptions of the service’s capabilities;
• instead of fixed call scenarios—dynamic construction of action chains;
• instead of a developer-centric approach—a model-centric approach, with the model as the primary consumer of the interface.

This shift affects not only the API format but also the overall system architecture: new requirements emerge for security, observability, and control over exactly what an agent does once it gains access to the tools.

The author of this article—a backend developer at Avito working at the intersection of traditional development and machine learning—emphasizes that the topic of integrating AI agents into business processes requires a more technical perspective, without excessive hype. The first part of the article is devoted to comparing the REST and MCP principles, as well as explaining why traditional approaches to API design need to be rethought with the advent of LLMs.

Source: All Articles in a Row / Machine Learning / Habr