festo-codesys-mcp

An MCP server that turns AI assistants into competent Festo/CODESYS PLC engineers.

LLMs are strong general programmers but unreliable PLC engineers: they invent Function Block signatures, hand-write PLCopen XML that CODESYS rejects, and guess at error codes. This server fixes that by grounding the model in curated, manufacturer-accurate knowledge and forcing every artifact through a machine-checked validation pipeline — so what reaches the CODESYS IDE actually imports, compiles, and follows one consistent engineering standard.

IEC 61131-3 ST  ──►  conventions enforced  ──►  PLCopen TC6 XML  ──►  3-stage validation  ──►  CODESYS import
     ▲                        ▲                                              │
     │                        │                                              ▼
  38-block library    27 knowledge topics + 19 manuals              official TC6 v2.00 XSD

Highlights

• 18 core tools — lookup, knowledge search (BM25), library reuse, ST/DUT/GVL generation, static analysis, convention review, PLCopen XML generation and a three-stage validation chain, ST code intelligence (symbols & where-used).

• Embedded knowledge base — 27 curated topics (Festo PtP motion, EtherCAT CiA 402, CPX-E, VTUX, CMMT-ST, MQTT/IIoT, PLCopen ground truth, naming conventions…) plus 19 reference manuals, all searchable full-text.

• Teaching modules that go beyond datasheets: PLC program architecture (Input-Process-Output, status-bus coordination), alarm design patterns, testing ST without hardware (digital-twin methodology), CODESYS Recipe Manager, hard-won CODESYS gotchas, and a complete custom web HMI + OPC-UA gateway curriculum (three-layer architecture, embedded-panel deployment).

• 38-block reusable ST library — 30 vendor-neutral Function Blocks plus 8 shared DUTs (motors, valves, PID, sensors, safety, logging, MQTT), all following one strict interface convention.

• Built for AI agents — always-on server instructions, prescriptive USE WHEN tool descriptions, next-step hints appended to every tool result, and an engineering-discipline topic that encodes non-negotiable quality gates (verify before delivering, no regressions, no hardcoded secrets).

• Optional live surfaces — 12 ide_* tools drive a running CODESYS V3.5 IDE (create/patch POUs, compile, browse the device tree), and 8 hardware tools commission Festo CPX-AP I/O and CMMT drives on the bench.

• Trustworthy by construction — ~690 automated tests (676 node:test + 15 pytest); the generator's output is validated against the official PLCopen XSD (v2.01 release, namespace-patched to the tc6_0200 namespace CODESYS emits).

Related MCP server: TwinCAT Validator MCP Server

Quick Start

Requires Node.js ≥ 20. No API keys, no accounts, fully offline.

Claude Code — add to .mcp.json in your project:

{
  "mcpServers": {
    "festo-codesys-mcp": {
      "command": "npx",
      "args": ["-y", "festo-codesys-mcp"]
    }
  }
}

Claude Desktop — add the same block to claude_desktop_config.json. On Windows, wrap the command with cmd /c:

{
  "mcpServers": {
    "festo-codesys-mcp": {
      "command": "cmd",
      "args": ["/c", "npx", "-y", "festo-codesys-mcp"]
    }
  }
}

From source:

git clone https://github.com/efranceschetti/festo-codesys-mcp.git
cd festo-codesys-mcp
npm ci && npm run build && npm test
# point your MCP client at: node build/index.js

Then just ask your assistant:

"Create a function block for a conveyor with jam detection, export it as PLCopen XML and validate it for CODESYS import."

The server steers the model through the full pipeline: load conventions → check the library for an existing block → generate → static-analyse → review naming → generate XML → validate (heuristics → XSD → semantics) → hand over an import-ready file.

Target Platform

Much of the knowledge (IEC 61131-3 conventions, PLCopen XML, CiA 402, program architecture, alarm patterns, twin testing) applies to any CODESYS V3.5 target, not only Festo hardware.

Tools

Core (18 — always available)

CODESYS IDE driving (12 — optional, Windows)

Enabled by setting FESTO_MCP_CODESYS_PATH to your CODESYS.exe and FESTO_MCP_CODESYS_PROFILE to the profile name shown in the CODESYS version selector. Drives a live CODESYS V3.5 instance through its scripting engine: open/create projects, create POUs/methods/properties, read and atomically patch POU code, compile and surface errors, and walk the project & hardware device tree as JSON. Use it after the offline pipeline produced validated XML — IDE changes are irreversible.

Hardware commissioning (8 — optional, bench only)

Enabled by FESTO_MCP_ENABLE_HARDWARE=1 (plus Python with the official festo-cpx-io / festo-edcon packages). Discover CPX-AP modules, read/write I/O channels, control CMMT drives (enable, ack faults, position tasks) and read/write raw PNU/SDO parameters. These tools bypass the PLC and write directly to devices — never use them on a machine in production.

Prompts (10 guided workflows)

new-function-block · new-motion-fb · new-project · new-ethercat-slave · debug-axis · decode-error-code · audit-naming · validate-st-batch · audit-plcopen-xml · convert-manual

Knowledge Base

Plus 19 auto-discovered manuals (CPX-E system/IO/EtherCAT/Profinet, CMMT servo & CiA 402 protocol reference, motion library, module catalog, CoE dictionary, data logging/FTP, OPC-DA, safety relay, VFD reference…). Drop your own converted manuals into knowledge/manuals/ and they are indexed automatically at runtime — no rebuild needed.

The web HMI curriculum

A distinctive part of this knowledge base: a complete, generic curriculum for building a custom web HMI for CODESYS PLCs — the three-layer architecture (SPA ↔ WebSocket ↔ Python gateway ↔ OPC-UA ↔ PLC), a tag manifest as single source of truth with five access modes (read/write/pulse/dead-man hold/ heartbeat), hosting the SPA on the PLC's own web server, and deploying to resource-constrained embedded panels (Chromium 69 on ARMv7) including the unified panel-installer pattern. Ask your assistant "how do I show PLC data in a browser?" and it will find its way here.

Built for AI agents

This server is designed so the model cannot quietly go off the rails:

• Server instructions (always in the client's context) pin the mandatory workflow — lookup → learn → reuse → create → verify → export — and five hard NEVERs (never invent FB signatures, never hand-write PLCopen XML, …).

• Prescriptive tool descriptions (USE WHEN / ALWAYS BEFORE / DO NOT) make the right next call obvious.

• Next-step hints: every successful tool result ends with a one-line Next: pointing to the next stage of the pipeline.

• engineering-discipline topic: non-negotiable quality gates — verify before delivering, re-validate the whole affected set after any edit, zero deferred warnings, no hardcoded credentials or network addresses, and human review before anything runs on a machine.

• For Claude Code users, the repo ships ready-made skills (ST writing, PLCopen pipeline, error diagnosis, motion control, HMI/gateway), review subagents, and a project-memory scaffold under .claude/.

Configuration

Development

npm ci
npm run build        # tsc → build/
npm test             # node:test
npm run lint         # eslint, zero warnings
npm run inspect      # MCP Inspector against the built server

cd python            # PLCopen XSD + semantic validation gates
uv sync && uv run pytest -q

The synthetic test fixture (python/tests/fixtures/synthetic-project-valid.xml, ~200 KB, 29 POUs) is generated by the server's own pipeline (scripts/gen-synthetic-fixture.mjs) and must always validate against the official TC6 v2.00 schema — regenerate it after generator changes and re-run the XSD gate to catch regressions.

See CONTRIBUTING.md for style rules and known future work (MCP SDK v2 migration, PLCopen ObjectId round-trip identity).

Safety notice

This project generates logic that can command real machinery. It is a reference implementation and engineering assistant — generated code must be reviewed by a qualified automation engineer before it runs on any machine that can cause physical harm, and safety functions must be implemented and validated per the applicable standards. See SECURITY.md.

Acknowledgments

This project stands on the shoulders of excellent open work, and is grateful to:

• codesys-mcp-toolkit by Johannes Pettersson (MIT) — the foundation for the ide_* CODESYS Scripting-Engine driving tools and their script templates.

• Festo SE & Co. KG — the official festo-cpx-io and festo-edcon Python SDKs, which the cpx_* / edcon_* commissioning tools wrap.

• PLCopen — the TC6 XML interchange standard and the reference schema/example that the generator and validators are built against.

• Model Context Protocol and the @modelcontextprotocol/sdk — the protocol and TypeScript SDK this server is built on.

Full third-party copyright and license notices are preserved in NOTICE.md.

License & attribution

MIT — © 2026 E3 Engenharia Industrial · Eduardo Franceschetti.

Built from real-world Festo/CODESYS machine engineering practice.

Festo® and CODESYS® are trademarks of their respective owners. This is an independent open-source project, not affiliated with or endorsed by Festo SE & Co. KG or CODESYS GmbH.