ThingWorx 10.2 brings agentic AI and MCP support to industrial IoT

ThingWorx, the industrial IoT platform PTC acquired in 2013, has spent more than a decade connecting factory machines to management systems, and it has now reached version 10.2. The release brings two headline additions: agentic AI capabilities and support for the Model Context Protocol (MCP). The news comes from ARC Advisory Group, a consultancy specialising in industrial automation, in a piece published this weekend. The article credits the release to Velotic, the company now behind the product.

This matters more than it might seem. MCP was born in November 2024 as Anthropic's standard for letting language models call external tools. In under two years it has moved from code editors and productivity tools into far more conservative territory: the software that sits between the PLCs and the management systems of a plant.

What happened

According to ARC's headline, ThingWorx 10.2 adds two pieces that are worth keeping apart. One is agentic AI, meaning agents that chain several steps over the platform's data instead of simply answering a query. The other is MCP, which can come in through two doors. The platform can act as a server and expose data and services to external clients such as Claude Desktop or Claude Code. It can also act as a client and consume third-party tools from its own agents. Before planning a pilot, the first thing we would check is which of the two this release covers, and to what extent.

Until now, connecting an LLM to ThingWorx meant going through its REST API and writing a fair amount of custom code. The platform's data model is organised into Things, Thing Templates and Thing Shapes, with properties, services and events. Much of the telemetry arrives from PLCs through connectors such as Kepware. That structure maps quite naturally onto MCP: a ThingWorx service looks a lot like a tool, and a property's history looks a lot like a resource.

Why it matters

Industrial IoT has spent years piling up data that almost nobody outside the engineering team ever queries. Say a maintenance manager wants to understand why line 3 lost throughput this week. Usually they have to open a dashboard, export to a spreadsheet and cross-check it against the MES. With a well-scoped MCP server, they can ask that question in natural language from the same client they already use for other tasks, and the model decides which services to call.

For integrators, the benefit is less flashy but more tangible: a standard connector replaces point-to-point integrations. If the protocol ships with the platform, the job stops being writing the bridge. It becomes deciding what gets exposed and with which permissions.

Governance of the standard itself also counts. MCP no longer depends on Anthropic alone: since December 2025 the Agentic AI Foundation, under the Linux Foundation, has been its steward. That reassures vendors that plan their platforms ten years ahead.

What to watch

Bringing agents into OT environments is not the same as connecting them to a task manager. Before switching anything on, we would run three checks:

1. Read versus write. Querying a sensor's history carries no operational risk; changing a setpoint does. The sensible approach is to start with read-only tools and require human approval for any action that touches the process.
2. Network segmentation. An MCP server with access to the control layer has to respect the zone-based architecture set out by standards such as IEC 62443, not bypass it for convenience.
3. Traceability. Every agent call should be logged with the user, the tool invoked and its parameters, just as a manual change in the SCADA is logged.

Who it is useful for

Mainly, plant engineering teams that already run ThingWorx and want to get more out of their data without building yet another dashboard. It also helps system integrators who until now sold custom connectors. And it helps data teams working with Claude or other MCP clients that were looking for a standard route into industrial telemetry.

At ElephantPink we like seeing MCP ship as standard in industrial platforms, because it removes a lot of glue code. But on the plant floor the question is not whether the agent can act, but who is accountable when it does, and that is still a design problem, not a protocol one.