"Read the factory temperature" -- Modbus MCP

正體中文

Look, nobody wants to memorize that register 40001 is the temperature sensor. Life is too short and Modbus addresses are too many. This is an MCP Server for Modbus TCP devices -- you define device profiles in YAML, and then AI agents can read/write PLC registers by name like civilized beings.

It also ships with a built-in simulator, because we all know you don't have a PLC sitting on your desk. (And if you do, please dust it off.)

---

Why This Exists

I kept looking at existing MCP servers for Modbus and kept finding the same four disappointments:

1. No semantic register mapping -- the AI has to know raw addresses like 40001. Asking it to "read temperature" gets you a blank stare.
2. No data type conversion -- everything comes back as raw uint16, because apparently float32 is too fancy.
3. No built-in simulator -- want to test? Go buy hardware. Or beg a colleague. Or stare at the ceiling.
4. No device profiles -- no way to pre-configure connection details and register maps, so you get to type the same host/port/slave_id over and over like it's 1998.

So I did the unreasonable thing and fixed all four myself.

---

Architecture

User (CLI / Chat / OpenClaw)
  → AI Agent (Claude / OpenClaw / etc.)
    → MCP Protocol (Streamable HTTP)
      → kc_modbus_mcp (FastMCP Server)
        → Profile Manager (YAML device profiles)
        → pymodbus (async Modbus TCP client)
      → Modbus TCP Device / Built-in Simulator

Features

• Natural language control -- "read the factory temperature" just works. No address lookup tables taped to your monitor.
• YAML device profiles -- map register names to addresses, data types, units, and scaling. Finally, configuration that reads like English.
• 8 MCP tools -- 5 profile-based + 3 raw mode for when you need to go full hacker.
• Auto data type conversion -- float32, int32, uint16, bool -- with byte order and scale support, because doing bit math in your head at 2 AM is not a personality trait.
• Built-in Modbus TCP simulator -- sine-wave temperature, random humidity/pressure, writable coils. Your imaginary factory is doing great.
• Docker-ready -- docker compose up -d and you're in business. Two containers, zero excuses.
• OpenClaw skill -- wrapper script for local LLM agents who prefer the command line lifestyle.

---

Demo

---

Security Notice: This is a POC/development project designed for trusted LAN environments. The MCP server does not implement authentication. Raw mode allows read/write access to any reachable Modbus device. Do not expose service ports to the public internet without additional security measures.

Quick Start

Three terminals, five minutes, zero soldering required.

1. Clone and install

git clone https://github.com/KerberosClaw/kc_modbus_mcp.git
cd kc_modbus_mcp
uv sync

2. Start the simulator

uv run python simulator.py
# Modbus TCP simulator running on port 5020

3. Start the MCP server

# In another terminal
uv run python server.py
# MCP server running on port 8765, loaded devices.yaml

4. Test with MCP client

npm install -g mcporter
mcporter config add modbus --url http://localhost:8765/mcp
mcporter call modbus.list_devices
mcporter call modbus.read_device device=factory_sensor register=temperature
mcporter call modbus.write_device device=factory_sensor register=motor_speed value=1500
mcporter call modbus.device_status device=factory_sensor

Or use Docker Compose

For the "I don't want to open three terminals" crowd (understandable):

docker compose up -d
# Simulator on :5020, MCP server on :8765

---

Device Profile (YAML)

This is where the magic happens. Well, "magic" is generous -- it's just YAML. But it beats memorizing hex addresses. Define your Modbus devices in devices.yaml:

devices:
  factory_sensor:
    host: 192.168.1.100
    port: 502
    slave_id: 1
    byte_order: big               # big | little | mixed
    registers:
      temperature:
        address: 0
        function_code: 3          # 3=holding, 4=input
        data_type: float32        # uint16, int16, uint32, int32, float32, bool
        scale: 0.1
        unit: "°C"
        access: read
        description: "Ambient temperature sensor"
      motor_speed:
        address: 4
        function_code: 3
        data_type: uint16
        unit: "RPM"
        access: read_write
        description: "Motor speed setpoint"
      pump_on:
        address: 0
        function_code: 1          # 1=coil
        data_type: bool
        access: read_write
        description: "Pump on/off switch"

Supported Data Types

Type	Registers	Range
bool	coil (1 bit)	true/false
uint16	1	0 – 65535
int16	1	-32768 – 32767
uint32	2	0 – 4294967295
int32	2	-2147483648 – 2147483647
float32	2	IEEE 754

Supported Function Codes

Code	Name	Access
1	Read Coils	read
2	Read Discrete Inputs	read
3	Read Holding Registers	read/write
4	Read Input Registers	read

---

MCP Tools

Profile Mode (primary)

The user-friendly stuff. The reason this project exists.

Tool	Description
list_devices	List all configured devices
list_registers	List all registers of a device with metadata
read_device	Read a named register -- returns converted value + unit
write_device	Write a value to a named register
device_status	Check if a device is online

Raw Mode (advanced)

For when you need to bypass all that nice abstraction and talk to registers like it's a bare metal Tuesday.

Tool	Description
read_registers	Raw read by host/port/slave_id/fc/address
write_registers	Raw write by host/port/slave_id/fc/address
scan_registers	Scan address range for non-zero values

---

Built-in Simulator

A pymodbus-based Modbus TCP server that pretends to be a factory. The data is fake but the protocol is real. No hardware needed -- your laptop is the factory now.

Register	Address	FC	Type	Behavior
temperature	HR 0-1	3	float32	Sine wave 20~30°C
humidity	HR 2-3	3	float32	Random 40~60%RH
motor_speed	HR 4	3	uint16	Holds written value
pressure	IR 0	4	uint16	Random 900~1100 kPa
pump_on	Coil 0	1	bool	Holds written value
valve_open	Coil 1	1	bool	Holds written value

---

Project Structure

kc_modbus_mcp/
├── server.py               # MCP Server entry point
├── simulator.py            # Built-in Modbus TCP simulator
├── devices.yaml            # Example device profile
├── src/
│   ├── profile.py          # YAML profile loader + register resolver
│   ├── client.py           # pymodbus async client wrapper
│   ├── converter.py        # Data type conversion (raw ↔ engineering value)
│   └── tools.py            # MCP tool definitions
├── openclaw-skill/         # OpenClaw skill wrapper
├── tests/                  # Automated tests
├── docker-compose.yml
├── Dockerfile
├── pyproject.toml
├── DESIGN.md               # Design document (Chinese)
├── .env.example
└── LICENSE

---

Environment Variables

Sensible defaults included, because nobody should have to configure things just to see if they work.

Variable	Default	Description
MODBUS_PROFILE	devices.yaml	Path to device profile YAML
MCP_HOST	0.0.0.0	MCP server bind address
MCP_PORT	8765	MCP server port
SIMULATOR_HOST	0.0.0.0	Simulator bind address
SIMULATOR_PORT	5020	Simulator port

---

OpenClaw Integration

For OpenClaw / local LLM agents, a wrapper script turns verbose MCP calls into something you can actually type without getting carpal tunnel:

modbus list
modbus status factory_sensor
modbus read factory_sensor temperature
modbus write factory_sensor motor_speed 1500

See openclaw-skill/SKILL.md for details.

---

TODO

Things I will definitely get to. Eventually. Probably.

• Multi-device connection pooling
• Polling mode with configurable cache interval
• Change detection (notify on value changes)
• Modbus RTU (Serial) support
• Web UI for device profile editing

---