MCP vs REST in Industrial AI: in the shopfloor, control matters more than flexibility.
In recent months, anyone working in Artificial Intelligence has heard about the Model Context Protocol (MCP), introduced by Anthropic with the goal of standardizing communication between LLMs and external resources such as companies’ databases, applications, data sources.
The Model Context Protocol (MCP) is based on a simple idea: to provide a language model with a set of tools and resources for accessing it, leaving it the freedom to decide what to use and when.
In practice, the system defines the features (tools, APIs, data), while the model decides how to orchestrate everything.
This approach works very well in exploratory, low-risk contexts with human supervision. For example, an AI assistant organizing emails can safely use MCP — any errors are verifiable and reversible.
At the core of MCP there is a strong assumption: an LLM can make reliable operational decisions. But an LLM remains, by its nature, a probabilistic model.
Outputs can appear correct without actually being so. LLM can change behavior in a non-deterministic way, and it does not guarantee operational consistency over time.
Let us imagine a complex production line: industrial furnaces, assembly stations, painting systems, hundreds of sensors, and data flows.
This information set is augmented with other data, such as production plans, operator shifts, scheduled maintenance, and unexpected events. The factory ecosystem generates what we can define as a “data puddle”: a heterogeneous mass of interconnected data.
In such a context, decisions must be traceable, processes must be deterministic, and safety is a priority. Entrusting operational logic to an LLM that autonomously decides the flow (as in MCP) introduces significant risks.
MCP defines what can be done (tools and resources) but does not define when, how, or in which sequence.
As a result, the model implicitly builds the operational plan. Even when the result seems correct, the path taken may be inefficient, not verifiable, and difficult to replicate. This is the real limitation in industrial environments.
The alternative: DAG + REST
At Alleantia (after testing a lot the MCP approach) we have adopted a different solution, based on:
|
Approach |
Who controls the process |
|
MCP |
The model |
|
DAG + REST |
The developer |
Interestingly, a DAG could include a node like: “do whatever you want, using these tools”. Which is, in fact, equivalent to MCP. This shows that MCP is a conceptual subset of a more structured system—but not the other way around.
Building agents based on LLMs implies a constant tension: we want to delegate decisions and repetitive tasks, but we also want results as reliable as those of a human expert. This requires controlled freedom, not total autonomy.
The Model Context Protocol represents a major step in the evolution of AI agents. However, in the context of the manufacturing industry, its limits are difficult to overcome.
An architecture based on DAG for orchestration and REST for communication offers today greater reliability, greater control, and better integration with existing systems.
The AI landscape is evolving rapidly, and these evaluations could change within a few months. But as of today, when safety, high costs, and operational continuity are at stake, control remains more important than flexibility.