MCP vs Traditional Integrations on ServiceNow

There are two ways to connect an AI assistant to ServiceNow. You can build a traditional integration, or you can use MCP. I spent a while doing the first one. Now I default to the second, and here’s why.

The DIY way

The do-it-yourself approach is familiar to every ServiceNow developer. You stand up a Scripted REST API, write the auth handling, parse the request, run a GlideRecord query, shape the response, and handle the errors. Then you do it again for the next use case. And again.

It works. But every integration reinvents the same plumbing: auth, pagination, error handling, rate limits. The logic that decides what the AI can touch ends up smeared across a dozen scripts. Six months later, nobody can answer “what can this assistant actually do” without reading all of them.

The MCP way

MCP, the Model Context Protocol, flips that. You run one server that exposes a deliberate set of tools and resources. The assistant calls those tools through a standard interface. You define the menu once, in one place.

The wins are concrete. You write one auth and error-handling layer instead of one per integration. You get a single readable list of what the assistant can do. Scopes and ACLs still enforce permissions on every call. And a new use case means adding a tool, not rebuilding the plumbing.

ServiceNow shipping a supported MCP server changed the math here. The standard plumbing is now something you configure and govern, not something you maintain.

When DIY still wins

MCP isn’t always the answer. If you’ve got a single, narrow, high-volume integration that will never change, a purpose-built Scripted REST API can be simpler and faster. DIY also wins when you need a response shape so specific that a general tool interface gets in the way.

The rule I use: if it’s one fixed pipe, build the pipe. If it’s a surface a model will explore and that will grow over time, use MCP.

My default

For anything AI-facing that will evolve, I reach for MCP first. It puts the governance in one place I can audit, and it stops every new idea from turning into another pile of integration glue.

The model will keep changing. The discipline of exposing a controlled surface and keeping your scopes honest will not.