Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Severstal Builds an Offline AI Assistant for Factory Automation Programmers

AI assistants for industrial automation (the control systems that run pumps and machine tools) have a flaw: Siemens' assistant works only with Siemens hardware. Severstal has built a prototype Copilot on a local LLM into its Flogic environment. It can tell you where the signal to PUMP_CTRL.START comes from, or add a 5-second delay before a pump starts. An edit is applied only after the engineer confirms it, and the code never leaves the network.

AI-processed from Habr AI; edited by Hamidun News
Severstal Builds an Offline AI Assistant for Factory Automation Programmers
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1Siemens Eigen Engineering Agent works only with TIA Portal and SIMATIC S7-1200/1500
Key metric #2CODESYS MCP Server sends project code to the provider when used with a cloud model
Key metric #3Copilot supports the IEC 61499 standard, and no comparable tool existed on the market

Factory automation programmers write the control logic for pumps, conveyors, and machine tools, and like ordinary developers, they want AI assistants. But off-the-shelf solutions are tied to a single vendor or send code to the cloud, which is unacceptable for a closed production site. The Severstal team that builds components for an open industrial control system (ICS, in Russian ASU TP, an automated process control system) has put together its own version: a prototype called Flogic Copilot, built into the SeverCode (Flogic) development environment.

It runs on a locally deployed large language model (LLM) and understands the IEC 61499 standard, under which a program is assembled from function blocks (FBs). It does not let data leave the corporate network. For example, an engineer can ask, "Where does the signal to PUMP_CTRL.

START come from?" or give the instruction, "Add a 5-second delay before the pump starts." If you build AI tools for industry, it is worth looking at how the team handled questions of trust and security.

What is already on the market

Siemens Eigen Engineering Agent connects to TIA Portal, picks up project context, and writes code in SCL and LAD. It also generates HMI visualizations and configures drives and equipment. The limitation is obvious: it works only with TIA Portal and SIMATIC S7-1200/1500 controllers. Alongside it sits Engineering Copilot TIA, which, according to Siemens' description, autonomously performs complex engineering tasks and modifies project elements.

CODESYS MCP Server is an official add-on component for CODESYS. It uses the Model Context Protocol (MCP), an open standard from Anthropic, to connect large language models to the development environment. There is a caveat: with a cloud model, all project data available to the user, including source code and library configurations, goes to the provider's servers. For confidential code, CODESYS itself recommends a local model.

For IEC 61499 projects built in Flogic, the authors say there was no comparable tool. Hence the decision: build a tool with Siemens-level features, but for this standard and with a local LLM.

What Flogic Copilot can do

The assistant sees only the project and its connected libraries, and does not reach out to external sources. It is invoked from the project's context menu. Its functions are as follows:

  • Project reference. Answers questions about the composition of applications, block connections, sources of events and data, and code in the ST language. It can also run a semantic search for blocks in the OSCAT and Standard Library libraries and in the project's own types: "Which block in OSCAT implements an off-delay?"
  • Verification. Checks the project using the environment's own tools: type mismatches, unconnected ports, parameter errors. You can ask why FB_TON is flagged as erroneous.
  • Block execution. Runs Simple FBs, Basic FBs, and functions in the Flogic interpreter on given inputs. The model does not guess the value of output Q for FB_TON with IN=TRUE for 3 seconds and PT=5 seconds. It gets the value from an actual computation.
  • Network editing. Adds blocks, sets parameters, and renames and deletes elements. An edit appears as a card with "Apply" and "Reject" buttons. In "Apply on my own" mode, the change is made immediately, followed by verification, and a rollback is possible.
  • New blocks. Writes blocks in ST, packages them as Simple FBs, and tests them in the environment. One example is a block that averages three analog inputs.
  • Algorithm editing. Changes the code of project types without touching the interface or the instances in networks. It shows the difference by running the code before and after the change.
  • OSCAT block porting. An engineer can paste a block description or code from the OSCAT documentation into the chat. The assistant then creates an FB, places it in the library, resolves dependencies, and adjusts the code to the environment's specifics.

Principles and limitations

The authors name three principles. First, the model receives a slice of the project graph (blocks, connections, types, parameters) and answers within its bounds. Second, test results come from executing the code in a deterministic environment, and the model only interprets them. Third, the decision stays with the human: changes are shown as a card, and rollback and version comparison are available.

The comparisons with Siemens deserve a sober reading. Claims such as "we are not inferior in context coverage" and "we are not inferior in the level of control over changes" are the team's own assessment, with no benchmarks or measurements. There is no data on which model runs inside, how accurate its answers are, or how much time an engineer saves. The plugin will be closed, not open source, even though the open software PLC itself is promised to be released as open source.

Planned features include: changing the interface of existing types and ECC Basic FBs, creating subapplications, configuring devices, resources, and mapping, uploading to the controller and debugging on FORTE, and running service and composite blocks such as E_DELAY and E_CYCLE.

The practical takeaway: customers with closed networks are waiting for local models, and trust comes from a scheme in which the model proposes, the environment checks and computes, and the human confirms. This combination can be adopted for other engineering environments as well.

Story

Key Questions Answered

What is already on the market?

Siemens Eigen Engineering Agent connects to TIA Portal, picks up project context, and writes code in SCL and LAD. It also generates HMI visualizations and configures drives and equipment. The limitation is obvious: it works only with TIA Portal and SIMATIC S7-1200/1500 controllers. Alongside it sits Engineering Copilot TIA, which, according to Siemens' description, autonomously performs complex engineering tasks and modifies project elements. CODESYS MCP Server is an official add-on component for CODESYS.

What Flogic Copilot can do?

The assistant sees only the project and its connected libraries, and does not reach out to external sources. It is invoked from the project's context menu. Its functions are as follows: Project reference. Answers questions about the composition of applications, block connections, sources of events and data, and code in the ST language.

What are the key details on Principles and limitations?

The authors name three principles. First, the model receives a slice of the project graph (blocks, connections, types, parameters) and answers within its bounds. Second, test results come from executing the code in a deterministic environment, and the model only interprets them. Third, the decision stays with the human: changes are shown as a card, and rollback and version comparison are available. The comparisons with Siemens deserve a sober reading.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…