node-red-contrib-mcp

node-red-contrib-mcp

The bridge between Node-RED and AI agents

Install · Quick Start · Nodes · AI Agent · Examples

Multi-phase OEE analysis agent built with MCP tool nodes in Node-RED

MCP (Model Context Protocol) is the open standard by Anthropic for connecting AI to external tools and data. This package brings MCP to Node-RED — the world's most popular low-code platform for industrial automation and IoT.

4M+ Node-RED installations meet 10,000+ MCP servers. Build AI agents visually. No code required.

Features

• Any MCP server — Streamable HTTP and SSE transport, with optional auth
• Any LLM — OpenAI, Anthropic, Ollama, vLLM, Azure, Gemini, or any OpenAI-compatible API
• AI Agent node — full agentic loop (tool discovery → LLM reasoning → tool execution → repeat)
• Zero lock-in — Apache-2.0 license, no cloud dependency, runs fully local
• Production-ready — error handling, status indicators, configurable timeouts
• Node-RED native — config nodes, msg passing, debug panel integration

Architecture

┌─────────────────────────────────────────────────────────────────┐
│  Node-RED                                                       │
│                                                                 │
│  [inject] → [mcp tool] → [llm call] → [mcp tool] → [debug]   │
│                                                                 │
│  [inject] → [ai agent] → [debug]    ← autonomous agent loop   │
│                                                                 │
└──────────┬──────────────────────────────────────┬───────────────┘
           │                                      │
           ▼                                      ▼
    ┌──────────────┐                      ┌──────────────┐
    │  MCP Server  │                      │   LLM API    │
    │  (any tool)  │                      │  (any model) │
    └──────────────┘                      └──────────────┘

Install

cd ~/.node-red
npm install node-red-contrib-mcp

Or search for node-red-contrib-mcp in the Palette Manager:

Menu → Manage palette → Install → node-red-contrib-mcp

Nodes

Node	Description
mcp server	Config — MCP server connection (URL, transport, API key)
llm config	Config — LLM provider (base URL, model, API key)
mcp tool	Call any MCP tool. Pass arguments as msg.payload, tool name in config or msg.topic
mcp tools	List all available tools from an MCP server. Great for discovery and debugging
mcp resource	Read resources exposed by an MCP server
llm call	Call any OpenAI-compatible LLM. Supports system prompt, JSON mode, multi-turn chat
ai agent	Autonomous agent — LLM + MCP tools in a reasoning loop until it has an answer

Quick Start

1. Call an MCP tool

[inject {"machine": "CNC-001"}] → [mcp tool "get_oee"] → [debug]

2. LLM + MCP pipeline

[inject] → [mcp tool "get_data"] → [llm call "Summarize this"] → [debug]

3. AI Agent (the magic node)

[inject "Why did OEE drop on machine 9014?"] → [ai agent] → [debug]

The agent autonomously discovers tools, reasons about which to call, executes them, and synthesizes a final answer. Same pattern as ChatGPT or Claude — but visual, auditable, and in your Node-RED.

AI Agent

The ai agent node runs a full agentic reasoning loop:

User: "Why did OEE drop on machine 9014 last week?"

  ┌─── Agent Loop ──────────────────────────────────────────┐
  │                                                         │
  │  Step 1: LLM sees 91 tools, picks get_oee              │
  │          → calls MCP server → gets OEE data             │
  │                                                         │
  │  Step 2: LLM analyzes, picks get_downtime_events        │
  │          → calls MCP server → gets 3 events             │
  │                                                         │
  │  Step 3: LLM synthesizes final answer                   │
  │                                                         │
  └─────────────────────────────────────────────────────────┘

Agent: "OEE dropped from 85% to 62% due to 3 unplanned stops:
        bearing failure (47min), tool change delay (23min),
        and material shortage (18min)."

  msg.agentLog = [{tool: "get_oee", ...}, {tool: "get_downtime_events", ...}]
  msg.iterations = 3

Agent settings

Setting	Default	Description
MCP Server	—	Which MCP server to use for tools
LLM	—	Which LLM provider for reasoning
System Prompt	—	Agent personality and instructions
Max Loops	10	Maximum LLM ↔ tool iterations
Temperature	0.3	LLM creativity (0 = focused, 1 = creative)

Examples

Import this flow

Copy the JSON below, then in Node-RED: Menu → Import → Paste

Example: MCP Tool Call

[
  {
    "id": "mcp-demo-inject",
    "type": "inject",
    "name": "Trigger",
    "props": [{ "p": "payload" }],
    "payload": "{\"machine_id\": \"CNC-001\"}",
    "payloadType": "json",
    "wires": [["mcp-demo-tool"]],
    "x": 150,
    "y": 100
  },
  {
    "id": "mcp-demo-tool",
    "type": "mcp-tool-call",
    "name": "Get OEE",
    "server": "mcp-demo-server",
    "toolName": "get_oee",
    "wires": [["mcp-demo-debug"]],
    "x": 350,
    "y": 100
  },
  {
    "id": "mcp-demo-debug",
    "type": "debug",
    "name": "Result",
    "active": true,
    "x": 550,
    "y": 100
  },
  {
    "id": "mcp-demo-server",
    "type": "mcp-server-config",
    "name": "My MCP Server",
    "url": "http://localhost:8021/mcp",
    "transportType": "http"
  }
]

Example: AI Agent

[
  {
    "id": "agent-demo-inject",
    "type": "inject",
    "name": "Ask question",
    "props": [{ "p": "payload" }],
    "payload": "What is the current OEE of machine CNC-001 and what are the main loss factors?",
    "payloadType": "str",
    "wires": [["agent-demo-agent"]],
    "x": 170,
    "y": 100
  },
  {
    "id": "agent-demo-agent",
    "type": "ai-agent",
    "name": "Factory Agent",
    "server": "agent-demo-mcp",
    "llmConfig": "agent-demo-llm",
    "systemPrompt": "You are a manufacturing AI assistant. Use the available MCP tools to answer questions about factory operations. Be precise and cite specific numbers.",
    "maxIterations": 10,
    "temperature": 0.3,
    "maxTokens": 4096,
    "wires": [["agent-demo-debug"]],
    "x": 400,
    "y": 100
  },
  {
    "id": "agent-demo-debug",
    "type": "debug",
    "name": "Agent Response",
    "active": true,
    "x": 620,
    "y": 100
  },
  {
    "id": "agent-demo-mcp",
    "type": "mcp-server-config",
    "name": "Factory MCP",
    "url": "http://localhost:8024/mcp",
    "transportType": "http"
  },
  {
    "id": "agent-demo-llm",
    "type": "llm-config",
    "name": "OpenAI",
    "baseUrl": "https://api.openai.com/v1",
    "model": "gpt-4o"
  }
]

Example: MQTT → AI Agent → MQTT (IIoT)

[
  {
    "id": "mqtt-in",
    "type": "mqtt in",
    "name": "machine/alerts",
    "topic": "machine/+/alert",
    "broker": "mqtt-broker",
    "wires": [["mqtt-agent"]],
    "x": 150,
    "y": 100
  },
  {
    "id": "mqtt-agent",
    "type": "ai-agent",
    "name": "Alert Agent",
    "server": "mqtt-mcp-server",
    "llmConfig": "mqtt-llm",
    "systemPrompt": "You are an industrial AI agent. When you receive a machine alert, investigate using MCP tools and recommend an action. Be concise.",
    "maxIterations": 5,
    "wires": [["mqtt-out"]],
    "x": 380,
    "y": 100
  },
  {
    "id": "mqtt-out",
    "type": "mqtt out",
    "name": "machine/actions",
    "topic": "machine/actions",
    "broker": "mqtt-broker",
    "x": 600,
    "y": 100
  }
]

MQTT alert comes in → AI agent investigates via MCP tools → action goes out via MQTT.

Compatible with

MCP Servers

Works with any MCP server that supports Streamable HTTP or SSE transport:

• OpenShopFloor — 91 manufacturing MCP tools (ERP, OEE, QMS, WMS)
• Anthropic MCP Servers — filesystem, GitHub, PostgreSQL, Slack, Google Drive, ...
• Any custom MCP server you build

LLM Providers

Works with any OpenAI-compatible API:

Provider	Base URL
OpenAI	https://api.openai.com/v1
Ollama (local)	http://localhost:11434/v1
Azure OpenAI	https://YOUR.openai.azure.com/openai/deployments/YOUR_DEPLOYMENT/v1
vLLM	http://localhost:8000/v1
LiteLLM	http://localhost:4000/v1
LM Studio	http://localhost:1234/v1
Anthropic	via LiteLLM proxy

msg Reference

mcp-tool-call

Direction	Property	Type	Description
Input	msg.payload	object	Tool arguments
Input	msg.topic	string	Tool name (if not set in config)
Output	msg.payload	any	Tool result (auto-parsed JSON)
Output	msg.mcpResult	object	Raw MCP response

ai-agent

Direction	Property	Type	Description
Input	msg.payload	string	User question or task
Output	msg.payload	string	Agent's final answer
Output	msg.agentLog	array	[{tool, args, result}] for each call
Output	msg.iterations	number	Total LLM reasoning steps

llm-call

Direction	Property	Type	Description
Input	msg.payload	string	User message
Input	msg.messages	array	Previous conversation (multi-turn)
Input	msg.tools	array	OpenAI-format tool definitions
Output	msg.payload	string	LLM response text
Output	msg.toolCalls	array	Tool calls (if any)
Output	msg.usage	object	Token usage stats

Configuration

MCP Server (config node)

Field	Description	Example
URL	MCP server endpoint	http://localhost:3001/mcp
Transport	Protocol variant	Streamable HTTP (default) or SSE
API Key	Optional Bearer token	sk-...

LLM Provider (config node)

Field	Description	Example
Base URL	OpenAI-compatible endpoint	https://api.openai.com/v1
Model	Model identifier	gpt-4o
API Key	Your API key	sk-...

Use Cases

Domain	What you can build
Manufacturing	OEE monitoring, capacity planning, predictive maintenance, quality root cause analysis
IIoT	MQTT → AI Agent → MQTT pipelines, sensor data analysis, anomaly detection
Building Automation	Smart energy management, BACnet/Modbus + AI reasoning
IT / DevOps	Database agents, log analysis, automated incident response
Prototyping	Fastest way to prototype agentic AI — visual debugging in Node-RED

Requirements

• Node-RED >= 3.0.0
• Node.js >= 18.0.0
• An MCP server to connect to
• An LLM API key (for llm-call and ai-agent nodes)

Contributing

Issues and PRs welcome! github.com/BavarianAnalyst/node-red-contrib-mcp

License

Apache-2.0 — use it anywhere, commercially or not.

Built by OpenShopFloor — the open-source AI platform for factory operations

Live Demo · GitHub · Node-RED Flow Library