Siemens Opcenter · MES / MOM

Opcenter integration that lets Claude read genealogy and WIP without touching a command

For manufacturers running Siemens Opcenter Execution who want an agent to answer MES questions from the system of record, under each person's own rights.

Opcenter integration for Claude means an MCP server in front of the read-only OData reading model of Siemens Opcenter Execution: named operations for genealogy, work in progress and non-conformances, run as the person asking and logged. On 19 September 2026 Siemens shipped no Opcenter MCP server. Pexon builds one, and transactional commands stay with people.

Why Opcenter is hard for agents

Four things that stop a generic connector at the MES

01 · Read path

The reading model is read-only, and that is the good news

Execution Foundation-based Opcenter products expose data for reading through a reading model over OData, with metadata a server can be built against. It is the safe path for an agent. It also only shows what your configuration projects into it, so a field that was never modelled is not there to read.

02 · Write path

Commands carry the plant's business logic

Changing state in Opcenter goes through commands that run business rules, and those commands are built per project. There is no generic write tool that works across installations, and a server that tried to offer one would be bypassing the logic the MES exists to enforce.

03 · Documentation

The developer guides sit behind Siemens' support portal

The Execution Foundation development and configuration guides are distributed through the Siemens Support Center under licence, not published openly. Anyone building a server needs your licensed access to them, and your MES team's knowledge of what was customised.

04 · Scope

Opcenter is a family, not one product

Execution Discrete, Execution Process and Execution Pharma, Opcenter X, Opcenter Intelligence and Opcenter APS have different interfaces. Opcenter Intelligence, for instance, has a documented REST API for KPIs on developer.siemens.com. An "Opcenter MCP server" is therefore built per product and version, never once for all of them.

  • Consequence: the first server reads, and it reads one Opcenter product. Writes are added one named command at a time, if at all.

Checked 19 September 2026

What Siemens ships for agents today, and what it reaches

What existsWhat it reachesWhat it does not do
MCP how-to on developer.siemens.comSiemens product documentation through llms.txt, using the open-source langchain-ai/mcpdoc serverReads no Opcenter data. It makes a coding assistant aware of API docs
Opcenter blog on AI agents (23 April 2026)Describes agents that plan, check and act, with Insights Hub as the analytics layer and Opcenter as the execution backboneDoes not mention the Model Context Protocol or a server a customer can deploy
Opcenter Intelligence REST APIKPI definitions and KPI results, scoped to the caller's tenant and rolesCovers analytics, not genealogy, work in progress or non-conformance records in Opcenter Execution
Community MCP servers on GitHubOne for Teamcenter (goenninger-b-t/tc-mcp)None for Opcenter was found in the repository search

A dated observation, not a permanent fact. Siemens and its partners are moving; the MCP connectors catalogue re-checks this row quarterly.

The first tool list

Seven operations a MES owner can read and sign

This is the shape of the manifest we agree before any code is written. Every tool reads, or drafts something a person submits. The entity sets behind each tool are confirmed against your reading model in the Blueprint, not assumed.

# opcenter-mcp: first tool list (read + draft only)
# Runs as the requesting user where the product allows it; every call logged.
tools:
  get_order_status:        # order, operations, current step, holds
    source: reading model (OData)
  trace_genealogy:         # "which serial numbers contain lot X?"
    source: reading model (OData)
    limit: max depth and result count set per plant
  list_wip_by_line:        # work in progress for one line or work centre
    source: reading model (OData)
  get_nonconformance:      # NCR record plus the operation history around it
    source: reading model (OData) or the quality module, per installation
  shift_handover_summary:  # open orders, holds, downtime notes for one shift
    source: composed from the tools above, no new access
  draft_ncr_disposition:   # proposal text only; a person submits it
    writes: nothing
  draft_shift_note:        # proposal text only; a person posts it
    writes: nothing
# Not on the list: any command. A write tool is added as one named,
# reviewed command, or not at all.

Two of the seven write nothing at all. That is the line a security review reads first, and the reason the list can be signed before the build starts.

What we build

The server, the identity, the log

  1. 01

    An MCP server against one Opcenter product

    Built against the reading model of the Opcenter product and version you run, with the tool list above adjusted to your configuration. Typed inputs, bounded results, no free OData query passed through from the model. Other systems join later as their own servers: the ERP order, the Teamcenter revision, the historian trend.

    • Model Context Protocol
    • OData reading model
    • typed tools
  2. 02

    The person's identity, as far as Opcenter can carry it

    Where your installation can authorise per user, the call runs as the person asking. Where it can only authenticate a technical integration user, we narrow the tools to what every permitted user may see, and record the requesting person on our side. Which of the two applies is tested in the Blueprint and written down before the build.

    • Entra ID or your IdP
    • per-user where possible
    • documented fallback
  3. 03

    Audit log, tests, and your repository

    Every tool call is logged with the person, the tool and the arguments. A test suite runs the tool list against a fixture of your data. The code, the tests and the runbook live in your repository and deploy into your network, next to the MES rather than in someone else's cloud.

    • audit log
    • eval fixtures
    • runbook
    • on-premises or your tenant

Where this is the wrong buy

If what you need is historised MES data joined to quality results for root-cause analysis across months, that is a data pipeline, not an agent tool, and a server that queries the reading model live is the slower way to get it. And if your Opcenter team is mid-upgrade, wait: a server built against a reading model that is about to change is rebuilt twice.

Start with the Readiness Blueprint.

Two weeks, €4,900 fixed price. We read your reading model with your MES team, agree the tool list with the MES owner, test whose identity each call can run under, and hand back a costed build plan for the first server. You keep it whoever builds the server. All prices are net and exclude VAT.

What we need from you: licensed access to the Opcenter developer documentation, a test or staging instance, and two half-days of your MES key user.

Related

Questions about connecting Opcenter to an agent

Siemens is investing in AI agents. Should we wait for an Opcenter MCP server?

Wait if Siemens announces one for your Opcenter product and version with a date. On 19 September 2026 there was none: the MCP how-to on developer.siemens.com serves product documentation, and Siemens' April 2026 Opcenter post on AI agents does not mention MCP. Even a vendor server would stop at Opcenter's boundary, so a question that needs the ERP order or the Teamcenter revision next to the genealogy still needs a server of your own.

Can the agent write to Opcenter, for example start or complete an operation?

Not in the first build, and possibly never. Writes in Opcenter Execution go through commands that carry the plant's business logic, and those commands are specific to each project's configuration. An agent can draft a non-conformance disposition or a shift note for a person to submit. A write tool is added only as one named command, reviewed with the MES owner, run under the requesting user and logged.

We run Opcenter Execution Pharma. Does GxP rule this out?

No, but it narrows it. In a validated MES the agent reads and proposes; electronic signatures are not delegable, and EU GMP Annex 11 expects batch certification to be done by a Qualified Person using an electronic signature. The server itself becomes part of your validated landscape, so its tool list, its tests and its audit log are written to be reviewed by your quality unit, which owns the validation.

What does an Opcenter integration cost?

The two-week Readiness Blueprint is fixed at €4,900: your Opcenter product and version, an agreed tool list, a test of whose identity each call can run under, and a costed build plan. A first production server is usually a use-case pilot from €15,000. Siemens licensing for API or integration users is between you and Siemens; check it before the build. All prices are net and exclude VAT.

What can this not do?

It cannot make an Opcenter installation expose data it does not already expose. If your reading model does not project a field, the server cannot read it without a change on the Opcenter side, which is your MES team's project, not ours. We also have no prior Opcenter build to reuse: every instance is configured differently, and the Blueprint exists to find those limits before anyone writes code.

Next step

Not a sales call. An architecture call.

Thirty minutes with the engineer who would read your reading model, on what an Opcenter server can and cannot expose in your installation.