Skip to content

IMTS 2026 Shows Industrial AI Adding Knowledge Graphs, MCP and Machine Hardware Links

Published
Author
Knud Lasse Lueth
Reads
1,000

IMTS 2026 put knowledge graphs, MCP servers and robot control interfaces in focus as vendors build an AI layer around industrial systems.

IMTS 2026 Shows Industrial AI Adding Knowledge Graphs, MCP and Machine Hardware Links

Industrial AI is beginning to take shape as a stack around the software and machines manufacturers already use. At IMTS 2026 in Chicago, the clearest signs were knowledge graphs that connect operational data, Model Context Protocol (MCP) servers that let agents query software in place, and a proposed interface for AI agents to discover and control physical equipment. The systems of record remain; the new question is how agents can use their data and capabilities with enough context and human oversight.

The International Manufacturing Technology Show ran September 14–19 at McCormick Place. Traditionally focused on CNC machine tools, additive manufacturing and fabrication, the 2026 event gave more floor space and attention to cloud providers, AI companies and industrial software vendors. The Association for Manufacturing Technology (AMT), the show’s organizer, introduced a dedicated Industrial AI Arena. Google presented its factory AI offering, while OpenAI showed an agentic maintenance system.

That shift matters because manufacturing AI depends on more than a model that can generate answers. An agent needs access to plant information, a way to interpret relationships between systems, and permission to take actions within defined limits. The technologies on display point to a developing architecture for those jobs, though the demonstrations and early deployments leave plenty of questions about reliability, integration and appropriate autonomy.

A new layer around ISA-95

The ISA-95 automation pyramid, in use since the 1990s, describes a familiar hierarchy: sensors and programmable logic controllers at the base, then supervisory control and data acquisition (SCADA), manufacturing execution systems (MES) and historians, with enterprise resource planning (ERP) at the top. Interviews at IMTS suggest that manufacturers and vendors are adding AI-oriented layers around this established structure rather than replacing its core records.

The emerging stack includes a software-defined edge, industrial DataOps, knowledge graphs and ontologies, interfaces such as MCP, agent-to-agent (A2A) links and the Model Hardware Standard (MHS), AI agents, and applications that keep people involved. These layers are still taking shape. Their purpose is to connect operational information to software that can reason over it and, where allowed, act on it.

One notable change is the role of the knowledge graph. For years, cloud strategies often started by copying plant data into a centralized data lake for analysis. AWS described a different approach at IMTS: its Industrial Data Fabric uses knowledge graphs, including Amazon Neptune, to connect and contextualize information where it resides across MES, SAP, product lifecycle management (PLM) and operational technology historians.

AWS also demonstrated an orchestrator that queries MCP servers connected to ERP, MES, computerized maintenance management systems (CMMS), PLM and its SiteWise historian. It then builds a knowledge graph on demand. The aim is to let an agent work with linked information without first consolidating every record in one repository.

Steve Blackwell, AWS head of product engineering and services for its Center of Excellence, said the knowledge graph had been part of the Industrial Data Fabric from its start, providing relationships and context for data from systems such as MES and SAP. He said Amazon Bedrock Agent Core adds GraphRAG, allowing an agent to work directly with the graph to reason across connected data.

For MachineMetrics, context is also a competitive asset. The company began as a machine-monitoring provider 12 years ago and now positions its MaximaOS platform as an MES. Chief revenue officer Graham Immerman described the company’s knowledge graph as its “moat,” arguing that data models and context matter more than a front end that others could build. That view captures a broader shift: as AI interfaces become easier to create, the quality of the underlying industrial model may determine whether an agent can return useful answers.

Trust depends on visibility

Access to data does not automatically make an agent trustworthy. Engineers need to see what an agent did, where its information came from and how certain it is before relying on a recommendation or action. AWS showed an Engineering Assistant that displays each reasoning step, the systems queried, the sources used and confidence scores. The design responds to engineers’ reluctance to assume an agent is right.

That kind of observability is a practical requirement for adoption. If a maintenance recommendation draws on a historian, a work order system and an equipment record, people need a way to inspect those inputs and understand the conclusion. Human review remains part of the picture at IMTS, particularly for actions that affect production equipment or maintenance work.

MCP connects agents to software in place

MCP offers a route for agents to query applications where data already lives. Anthropic introduced the open protocol in November 2024; the Linux Foundation now oversees it. Under the older copy-first pattern, manufacturers used extract, transform and load (ETL) jobs to move data from separate systems into a data lake before analysis. MCP servers can instead expose selected system capabilities to an agent at query time.

Several industrial vendors showed MCP servers at the event, including MachineMetrics, PTC’s Onshape, Rockwell Automation’s Emulate3D, Limble and Vention. The interfaces vary in what they allow. A native server can expose an application’s logic so an agent may change or run something inside it. A wrapper around a closed API may offer read-only access, while a server that simply transfers data between systems may lack the context needed for more meaningful work.

Vention chief product officer Brendan Sterne described using the company’s MCP with Claude Code to retrieve a design’s geometry and configuration, program it, then upload and run it on a digital twin. The example points to a broader use: an agent can move between design information and a simulation environment through a defined software interface. How much it can change or execute depends on the specific server and controls.

MHS extends the interface to physical equipment

Anthropic’s Model Hardware Standard aims to give AI agents a common way to discover and operate physical equipment, including robot arms, sensors and test instruments. Released as a research preview on August 27, 2026, MHS describes each device through a standard driver with “read” and “write” commands, a plain-language account of its capabilities and stated safety limits. Anthropic says the approach could cut integration work from weeks or months to hours or minutes.

That time saving is a vendor claim, not a reported comparative test result. Early MHS partners are mostly laboratories, while Universal Robots, Doosan Robotics and AWS are among the companies adding support. The concept is similar to MCP at the software level: agents discover available capabilities through a shared interface instead of requiring bespoke integration for each device.

Universal Robots tested MHS with four cobots through PolyScope X, its API-first control system. The company built an agent using Claude Opus 4.8 in Claude Code. Given plain-language instructions, the agent discovered the robots and their capabilities, configured them, and ran them as a cell, passing payloads between robots within each device’s declared limits.

Universal Robots called the setup a proof of concept and provided no performance figures. It did not demonstrate MHS on the show floor, though it discussed the potential with attendees. Chris Savoia, the company’s head of ecosystem, likened it to MCP for connecting hardware and said it enables natural-language tasking and native robot provisioning through Claude. Vention described MHS as an “introspection and control framework” that tells an AI model which sensors and actuators are available.

Vendors are choosing different speeds

The show also exposed a gap between companies exploring physical AI and those presenting a more developed platform. At Automatica 2025, Fanuc declined to discuss physical AI. At IMTS 2026, it showed two demonstrations with Google and AWS, both still in research and development. Universal Robots, by comparison, launched its Gen 7 cobot at IMTS as an open platform for AI applications and described its MHS test.

Not every vendor wants to give agents the same freedom. MaintainX cofounder Nick Haase said the company’s goal was to be “a half step behind.” That caution reflects a key design choice for industrial AI: how far a system can act without a person approving the next step. The answer will vary with the task, the equipment and the consequences of an error.

Across maintenance and asset performance, software vendors are also working toward tighter links between CMMS workflows and predictive maintenance. Connecting a sensor alert to a completed repair could let software coordinate more of the process. For now, the direction is toward connected loops with people still in control of important decisions.

IMTS 2026 therefore offered a view of the emerging industrial AI battlegrounds: knowledge graphs supply context, MCP opens software capabilities to agents, and MHS proposes a shared route to physical devices. Those pieces could make it easier to connect information and action across design, production and maintenance. The next test is whether vendors can make these links dependable and inspectable enough for engineering teams to use them in day-to-day operations.

Source: Knud Lasse Lueth · iot-analytics.com

Related reading