Command Palette

Search for a command to run...

UnylyUnyly
Весь каталог

Control Expert

БесплатноНе проверен

MCP server that enables AI agents to read, edit, and create PLC projects in EcoStruxure Control Expert via COM automation.

GitHubEmbed

Описание

MCP server that enables AI agents to read, edit, and create PLC projects in EcoStruxure Control Expert via COM automation.

README

An MCP server that connects AI agents (Claude Desktop, Claude Code, VS Code Copilot, or any MCP client) to a local EcoStruxure Control Expert (formerly Unity Pro) installation — so the AI can read, edit, and create PLC projects from scratch.

Inspired by tiaportal-mcp for Siemens TIA Portal. Where TIA Portal exposes the Openness API, Control Expert exposes the Unity Developer's Edition (UDE) COM automation server (PSBroker.PServerBroker) — this server drives it through pywin32 and exposes it as MCP tools over stdio.

AI agent (Claude, Copilot, ...) ──MCP/stdio──> control-expert-mcp ──COM──> Control Expert

What the AI can do

  • Projects: create from scratch for any CPU in the hardware catalog (M340, M580, Premium, Quantum...), open/save .stu/.sta, import/export .xef/.zef
  • Browse: tasks, sections, variables, DFB/DDT types, CPU/hardware info, build state
  • Program: create/read/delete sections in ST, LD, FBD, SFC, IL; write logic by importing Control Expert XML (the AI reads an existing section's XML once and mirrors the schema — validated round-trip for all five languages)
  • Hardware: walk the full bus → drop → rack → module topology, add/remove IO modules by catalog part number, change the CPU
  • Modbus / DTMs (validated live on M580): browse the DTM topology, add slave DTMs (add_dtm("Modbus Device", ..., parent_dtm="BMEP58_ECPU_EXT") — the protocol id is Modbus and is auto-tried), set device IP addresses, read the scanner state (get_dtm_control_parameters), and add Modbus scan requests (read/write addresses + sizes) by editing the master DTM dataset through get_master_dtm_dataset / set_master_dtm_dataset — a ManagedModbusRequest template with the numbering rules is built into the tool description; a written scan line builds built_ok and shows up as a ModbusScanLine in the scanner config
  • Networks (Premium/Quantum): create logical Ethernet networks and set their IP service configuration
  • Variables: list/create/update/delete global variables incl. type, comment, address (%MW...), initial value
  • Build: analyze and build the project, get the resulting build state
  • UI: pop the Control Expert window open so a human can watch or take over, and open an animation table's editor live (open_animation_table) to watch values animate while the AI drives a test
  • Online (opt-in): connect to PLC/simulator, download/upload, run/stop
  • Live values (opt-in): read and write running PLC/simulator values over Modbus TCP (modbus_connectread_tags/write_tags) — the UDE/COM API has no live tag access, so this is the channel for testing logic against a running controller (located %M/%MW tags; mirror unlocated DFB internals to %MW first)

Requirements

  • Windows with EcoStruxure Control Expert (or Unity Pro) installed and licensed — tested against Control Expert 14.0, but any version that registers PSBroker.PServerBroker.1 should work (the UDE automation server ships with Control Expert itself; the separate UDE package is only needed for documentation)
  • Python 3.10+ with pywin32
  • The server must run on the same machine as Control Expert (DCOM remoting is possible but not configured here)

Install

cd control-expert-mcp
python -m venv .venv
.venv\Scripts\pip install -e .

Hook up an AI client

The server speaks MCP over stdio. Point your client at the venv's Python:

Claude Desktop

%APPDATA%\Claude\claude_desktop_config.json (see samples/claude_desktop_config.json):

{
  "mcpServers": {
    "control-expert": {
      "command": "C:\\path\\to\\control-expert-mcp\\.venv\\Scripts\\python.exe",
      "args": ["-m", "control_expert_mcp"]
    }
  }
}

Claude Code

claude mcp add control-expert -- C:\path\to\control-expert-mcp\.venv\Scripts\python.exe -m control_expert_mcp

VS Code (GitHub Copilot agent mode)

.vscode/mcp.json (see samples/vscode-mcp.json):

{
  "servers": {
    "control-expert": {
      "type": "stdio",
      "command": "C:\\path\\to\\control-expert-mcp\\.venv\\Scripts\\python.exe",
      "args": ["-m", "control_expert_mcp"]
    }
  }
}

Tools

Tool Purpose
get_status Server version, open project, CPU, build state
open_project Open .stu / .sta / .xef / .zef
new_project Create a project for a CPU (e.g. BMX P34 2020 + 02.70)
save_project / close_project Persist / close (close discards unless save=true)
build_project / analyze_project Build or analyze; returns resulting build state
get_project_structure Tasks, sections (+language), events, fct modules
list_variables Variables with type/comment/address/initial value
create_variable / update_variable / delete_variable Variable editing
read_section Section logic as Control Expert XML
write_st_logic Write ST logic as plain IEC text (no XML) — creates or replaces a section, with inline variable declarations
get_language_reference Authoring guide + validated example for ST/LD/FBD/SFC/IL — the exchange-XML structure rules an AI client needs to write graphical logic (FBD pin-geometry rule, LD line/cell model, SFC chart layout)
validate_xml Pre-validate candidate exchange-XML against the installed SrcXmlSchema XSDs before import/build — instant structural errors with the legal value sets
place_fb_in_ladder Drop any project DFB into Ladder, no template — reads the DFB interface and generates the correct (CE-owned) pin geometry automatically, then bind pins
use_fb_in_ladder Clone a GUI-authored block-in-LD template and rebind instance + variables (fallback for elementary EFBs / exact GUI geometry)
create_section / delete_section Section management (ST/LD/FBD/SFC/IL)
create_task Add MAST/FAST/AUX/SAFE task
import_xml The main write path — import section logic, variables, DFB/DDT types, configuration, or whole-project exchange files (inline XML or file)
export_xml Export variables / program / configuration / one DFB / one section as XML
export_project Full application export to .xef / .zef
list_data_types DFB + DDT types
get_hardware CPU + bus → drop → rack → module tree
add_io_module / remove_io_module Add/remove rack modules by part number + catalog version (e.g. BMX DDI 1602 + 02.00)
add_drop / add_rack Build out remote drops/racks (e.g. an X80 EIO drop + BME XBP 1200 rack on an M580 RIO bus)
change_cpu Swap the CPU reference
list_animation_tables / create_animation_table / delete_animation_table Watch/animation tables (accepts hierarchical paths like Pump1.Speed, SFC_Demo.S_Init.x) for monitoring/forcing values online
list_dtms DTM Browser topology with names, types, addresses
add_dtm / delete_dtm Add communication DTMs or slave devices (e.g. Modbus Device under the M580 CPU DTM)
set_dtm_address Set a slave's IP / bus address (+ optional gateway/subnet — fixes the "IP Address and Gateway address are not in the same domain" build warning)
configure_cpu_ethernet M580 CPU embedded Ethernet: IP/subnet/gateway + Security-screen services (tftp/eip/dhcp_bootp/ftp/web/snmp). Enabling tftp+eip+dhcp_bootp clears the remote-EIO CRA build error
get_dtm_control_parameters / set_dtm_control_parameters Modbus TCP I/O scanner config XML (CetControlParameter* schema): IP, unit id, timeouts, ModbusTcpRequest scan lines
get_dtm_dataset / set_dtm_dataset A slave DTM's own dataset (identity + bus address)
get_master_dtm_dataset / set_master_dtm_dataset The master/CPU DTM dataset (via ZEF round-trip) — <SlaveDevices>/<ManagedModbusRequestList> holds the Modbus scan lines; write a ManagedModbusRequest node to add a request
list_networks / add_network / set_network_ip Logical networks on Premium/Quantum
show_ui Make the Control Expert window visible (read-only by default so it can follow along / open editors while the client keeps write)
open_animation_table Open an animation table's editor in the live CE window so a human watches values animate during a test
plc_* (opt-in) Online: setup/connect/disconnect/state/transfer/run/stop
modbus_connect / modbus_disconnect / modbus_status (opt-in) Open/close a Modbus TCP link to the CPU/simulator's server (127.0.0.1:502 for the sim) for live values
read_tags / write_tags (opt-in) Read/write LIVE located-tag values (%M/%MW) by name or address (%MW86, %MW2:REAL); the only live read/write path (COM API has none)

How the AI writes logic

Program sections are exchanged as Control Expert XML. An ST section looks like:

<STExchangeFile>
  <fileHeader company="Schneider Automation" product="Control Expert V14.0 - 190112"
              dateTime="date_and_time#2026-6-11-1:0:0" content="Structured source file"
              DTDVersion="41"></fileHeader>
  <contentHeader name="Project" version="0.0.000"></contentHeader>
  <program>
    <identProgram name="Logic01" type="section" task="MAST"></identProgram>
    <STSource>
IF StartButton AND NOT StopButton THEN
    MotorRun := TRUE;
ELSE
    MotorRun := FALSE;
END_IF;
    </STSource>
  </program>
</STExchangeFile>

import_xml(kind="section", xml_content=...) creates the section if it doesn't exist, or deletes-and-replaces it on conflict. The agent should declare referenced variables first (create_variable) or include a <dataBlock>, then build_project to validate.

Online tools (live PLC / simulator) — disabled by default

Tools that touch a controller (plc_connect, plc_transfer, plc_command, ...) are not registered unless you set the environment variable:

CE_MCP_ENABLE_ONLINE=1

Starting/stopping a PLC or downloading an application affects the physical process. Only enable this on test benches or with the simulator, and keep a human in the loop.

Simulator commissioning (validated end to end): launch the PLC simulator first — start_simulator or PLC_Simulator\sim.exe from the Control Expert install (it sits in the tray) — then plc_setup_connection("simulator", "127.0.0.1")plc_connect("simulator")plc_transfer("pc_to_plc")plc_command("run"). The project must build clean (built_ok) before download. plc_state reports connection + run/stop and whether PC == PLC.

Known limitation: a freshly started simulator with no station loaded (plc_state = no_conf) rejects plc_transfer with "Family check failed" — the API download (unlike the Control Expert GUI's) requires the sim to already have a station of a matching family. Seed it once by transferring any project from the Control Expert GUI (PLC → Simulation Mode → Connect → Transfer); the loaded station persists while sim.exe runs, and all API downloads work from then on (stop the PLC first — transfer to a running PLC fails).

Live values over Modbus TCP (validated end to end)

The UDE/COM automation API serves the project database — it has no live tag-value read/write (animation tables only render in the GUI; IVariable exposes only the offline initial value). Live values go over the CPU's Modbus TCP server instead — the same channel SCADA/Vijeo use:

modbus_connect(host="127.0.0.1")          # the sim's Modbus server; or a real CPU IP, port 502
read_tags("EMFM1FLOW, RESIDUAL2CLTHSP, %MW0, %MW2:REAL")
write_tags({"RESIDUAL2CLTHSP": 4.0, "%MX100.0": true})
  • Tags are global variable names (address + IEC type resolved from the project) or explicit addresses with optional :TYPE (%MW86, %MW2:REAL, %MW70:UDINT, %M3, %MW10.2 bit). Decoding is type-driven (INT/UINT, DINT/UDINT, REAL, BOOL coil/word-bit).
  • Only LOCATED %M/%MW tags are reachable. Unlocated DFB internals (e.g. Pump1.Running) must be mirrored to %MW/%M in the program first (see the test_logic_live prompt).
  • 32-bit REAL/DINT use Schneider low-word-first order by default; pass word_order="high_first" if a server differs.
  • The Control Expert simulator exposes a Modbus server on 127.0.0.1:502, so the whole read/write test loop works against the sim — no hardware required.

Prompts (guided workflows)

The server ships MCP prompts (src/control_expert_mcp/prompts.py) — reusable recipes that encode the validated flows and their non-obvious gotchas so a client doesn't rediscover them by trial and error. In Claude Code they appear as /mcp__control-expert__<name> slash commands; other MCP clients list them in their prompt picker.

Prompt What it walks you through
commission_simulator Build → start sim → connect → transfer → run, incl. the manual first-transfer / "Family check failed" seed step
test_logic_live The Modbus test loop for a DFB instance, incl. mirroring unlocated internals to %MW and word-order gotcha
author_logic The get_language_reference → write/import → build_project → fix-from-output loop (per language)
scaffold_project new_project (exact CPU+version) → rack/PSU/IO → build → save
add_modbus_device Add a Modbus-TCP slave DTM under the M580 CPU and configure a scan line

Extending the server

When you add a tool or capability, also add (or extend) an MCP prompt in prompts.py that walks a client through using it. This is a project convention, not an afterthought: a good prompt turns an hour of trial-and-error into one slash command.

Write a prompt for anything a client cannot guess from the tool description alone — an environment flag (CE_MCP_ENABLE_ONLINE), a manual GUI step (the simulator family-check seed), an ordering constraint (stop the PLC before transfer), or an addressing rule (located-only Modbus, low-word-first REALs). Rules of thumb:

  • Each new workflow → a prompt; each new tool → at least a mention in a relevant prompt.
  • Put the gotcha in the prompt text explicitly (the steps that cost you time while building it).
  • Mirror the validated sequence (tool names + argument shapes), and reference related prompts by name.
  • Keep the standing orientation in the server instructions string short; put step-by-step recipes in prompts.

Troubleshooting

  • Catalog object not found on new_project — the CPU part number/version must exactly match the hardware catalog of your Control Expert version (spacing matters: BMX P34 2020, firmware like 02.70 is required).
  • application object reference is not found in the catalog on add_io_module — same rule for modules: the catalog version is mandatory (02.00 for most M340 IO modules, 01.00 for racks/power supplies).
  • new_project fails with a bare Exception occurred for a CPU family — your Control Expert license/DTM library probably doesn't include that platform. The error carries no description; test the same CPU in the Control Expert GUI to confirm.
  • add_dtm says "protocol Id parameter is empty or invalid" / "Impossible to create a new DTM object" — slave DTMs need the FDT protocol name as protocol_id (Modbus for the generic Modbus TCP device; the server auto-tries Modbus and EtherNet/IP when empty) and a device_type_name that exactly matches the DTM catalog (Modbus Device). List the catalog names with tools/list_dtm_catalog.py.
  • set_master_dtm_dataset crashes the server (RPC failed) — almost always malformed request XML; in particular requestUniqueID must be a fresh GUID, not an integer. Mirror the validated template in the tool description exactly.
  • Write access mode is already reserved by another client — Control Expert (the GUI) or another automation client has the project open for writing. Close it there first. The server holds one write token per session by design.
  • Broker creation fails — Control Expert isn't installed, or its COM registration is broken (re-register by repairing the installation).
  • First call is slowopen_project/new_project start the Unity server process and can take 30–120 s for large projects. Configure generous tool timeouts in your client.
  • Import fails with Invalid file — the XML doesn't match the exchange schema. Export a similar object first (read_section, export_xml) and mirror its structure, including fileHeader/contentHeader.

Architecture notes

  • All COM calls run on a single dedicated STA worker thread (COM apartment affinity); MCP tool calls are marshalled onto it (the dispatcher is reentrant — nested bridge calls on the worker thread execute directly).
  • Hardware and DTM objects expose their members on secondary dual interfaces (IProject3, IConfiguration2, IBus, IModule, IPServerDtm*, ...) that the default dispinterface doesn't include. The bridge QIs each object with IIDs harvested from HKCR\Interface and wraps the result as IDispatch (_qi), which is the only way to reach DTMRoot, InternalBuses, AddChild, etc. from late-bound clients.
  • The write-access token (app.Project(1)) is acquired once per session and cached; failed-call tracebacks are stripped so they can't pin COM references and deadlock ProjectClose.
  • Project is a parameterized COM property — it is invoked with explicit DISPATCH_PROPERTYGET flags because pywin32 dynamic dispatch can't call it.
  • Import temp files get the extension matching the XML root element (STExchangeFile.xst, VariablesExchangeFile.xsy, ...) because Control Expert picks the parser from the extension.
  • Enum constants (languages, export options, PLC commands...) were extracted from PServer.tlb — see src/control_expert_mcp/constants.py.

Disclaimer

Not affiliated with Schneider Electric. The UDE automation interface is provided by Schneider "as is" and is no longer commercialized; this project drives it at your own risk. Never point online tools at production equipment without qualified supervision.

from github.com/apexsotjo-blip/control-expert-mcp

Установка Control Expert

У этого сервера нет опубликованного пакета — он собирается из исходников. Открой репозиторий и следуй инструкции в README.

▸ github.com/apexsotjo-blip/control-expert-mcp

FAQ

Control Expert MCP бесплатный?

Да, Control Expert MCP бесплатный — установка в пару кликов через Unyly без оплаты.

Нужен ли API-ключ для Control Expert?

Нет, Control Expert работает без API-ключей и переменных окружения.

Control Expert — hosted или self-hosted?

Self-hosted: сервер запускается локально на твоей машине командой из раздела установки.

Как установить Control Expert в Claude Desktop, Claude Code или Cursor?

Открой Control Expert на unyly.org, выбери вкладку своего клиента (Claude Desktop, Claude Code, Cursor) и нажми Install — конфиг сгенерируется автоматически, без правки JSON.

Похожие MCP

Compare Control Expert with

Не уверен что выбрать?

Найди свой стек за 60 секунд

Автор?

Embed-бейдж для README

Похожее

Все в категории ai