There’s a moment in every project when the team has to answer a familiar question: should we use an existing tool, or should we build our own? For our colleagues working with agentic workflow architecture, that question became clear with MCP, or Model Context Protocol.
What Is Model Context Protocol (MCP) and Why Do AI Agents Need It
Model Context Protocol is a standard that defines how AI agents communicate with external tools and data sources. Instead of each LLM integrating directly with every service, MCP introduces a uniform layer: a server exposes tools, a client discovers and calls them, and the agent decides what to use and when.
The protocol consists of three main steps: initialize, list tools, and call tools. Its simplicity makes it appealing for agentic workflows. The LLM gets a list of available tools along with their descriptions. It then decides, based on context, which one to use.
This idea sounds great in theory, and it is. But the real test comes when you start using it in practical situations.
Using N8N as an MCP Server: Built-In Support, Real-World Limits
N8N is a workflow automation platform that, in its newer versions, offers built-in MCP support. For the team, it was a logical first step – the platform was already part of the stack, the integration is relatively fast, and the visual editor makes prototyping straightforward.
What worked well:
- Fast setup and a lower entry barrier
- Visual workflow creation without writing custom code for every step
- Solid fit for standard, predictable flows
Limitations they ran into:
- Reduced flexibility when implementing non-standard tools and custom logic
- Dependency on what the platform exposes through its MCP interface
- Less control over endpoint definitions and server behavior
N8N MCP is a good fit for proof-of-concept work and fast iteration. But once a project starts demanding granular control, it gets tight.
Building a Custom MCP Server and Deploying It on AWS
The second path the team explored: a fully custom MCP server implementation, deployed on AWS infrastructure.
The architecture follows the MCP spec directly – the server exposes endpoints, implements list tools and call tools methods, defines input schemas for each tool, and the client connects and communicates over the protocol. Every tool has clearly defined descriptions – and that’s not cosmetic detail. Descriptions are what the LLM reads when deciding what to call. A weak description leads to a poor agent decision.
Advantages of the custom approach:
- Full control over what the server exposes and how
- Direct integration with a knowledge graph and custom services from code
- Flexibility in defining tools, parameters, and endpoint logic
- Adaptable to specific project and client requirements
Trade-offs:
- Higher initial development and maintenance overhead
- Requires a well-defined deployment and versioning process
- The team needs to own both the MCP layer and the business logic beneath it
How MCP Fits Into an Agentic Workflow Architecture
Both approaches live inside the broader agentic workflow concept – an architecture where an AI agent isn’t a passive instruction-executor, but an active participant that plans, selects tools, and adapts based on results.
The key point the team demonstrated: the MCP server is just one piece of the puzzle. Equally important is how the agent is configured, what tool descriptions it receives, and how the overall workflow is orchestrated. A ready-made server like N8N MCP can be sufficient when workflows are predictable. When agents need to make more complex decisions and call more specific logic — the custom approach delivers the precision required.
N8N MCP vs. Custom MCP Server: Which One Should You Use
This isn’t about one solution being objectively better than the other. It’s about understanding the limitations and strengths of each approach before you hit production – because that understanding is the difference between a project that runs and one that struggles.
N8N MCP is a valid tool for fast iteration and less complex integrations. A custom MCP server on AWS gives you full control where it genuinely matters.
Real expertise? Knowing which one to reach for.
In this article:
Levi9 Serbia
