Over the past few months the Model Context Protocol (MCP) has emerged as the first serious attempt to standardize how Large Language Models talk to external systems. For anyone working in PLM — and especially in Siemens Teamcenter — this is more interesting than it might look at first glance.

What MCP is, in two lines

MCP is an open protocol that defines how an LLM can discover and invoke tools exposed by an external system. It’s effectively the USB equivalent for AI agents: a common interface that decouples the model from specific integrations. Once a system “speaks MCP”, any compatible model can use it — a certain LLM today, a better one tomorrow, with no integration rewrites.

Why it makes sense on Teamcenter

Teamcenter is a heavy enterprise system: functionally rich, but with a user experience that reflects its desktop-client heritage. Structured queries (saved queries, attributes, relations) are powerful but require familiarity that most users who don’t work with it every day don’t have.

Exposing Teamcenter operations as MCP tools changes the interaction model:

  • A PLM Manager types “show me all the revisions released this month for the Alpha project, grouped by family” — and gets the answer, executed as a real query, not as a hallucination.
  • An approval workflow can be started conversationally: the agent confirms what it’s about to do and, via Function Calling, proceeds or stops based on the user’s reply.
  • Periodic reports get automated without writing Java reportlets: the LLM drafts the structure, the agent retrieves the data, the PDF file is ready.

Four technical aspects that make the difference

  1. Tool discovery. The model doesn’t need to “know” what Teamcenter can do: it discovers it at runtime from the MCP server. This lets you extend the exposed surface without retraining.
  2. Schema-driven invocations. Each tool has an explicit schema (inputs, outputs, errors). Calls are validated before hitting Teamcenter — no more malformed requests.
  3. Reversibility and dry-runs. Sensitive operations can be preceded by a preview executed by the agent, to approve before the apply. That’s the difference between demo and production.
  4. Multi-tenancy and auditing. Calls go through the user’s identity, not an omnipotent service account. What the agent does, it does on behalf of whoever’s chatting.

FITEC’s implementation: PaLoMa

This isn’t theory: it’s the foundation on which we built PaLoMa, our AI assistant for Teamcenter. PaLoMa orchestrates specialized agents (search, BOM, workflow, reporting) over a custom MCP server that talks to Teamcenter via API. The value isn’t in the LLM (interchangeable) — it’s in the MCP server, in the tools it exposes, and in the discipline with which they are written.

If you’re evaluating AI on top of your PLM, MCP is the right starting point — and it isn’t Teamcenter-specific. The same pattern works on ERP, MES, technical documentation systems.