Industrial AI Agents with Model Context Protocol (rekuiper-mcp)

Physical AI and industrial edge operations require AI agents to inspect telemetry streams, validate SQL transformations, and manage gateway pipelines. rekuiper-mcp is a native Rust server implementing the Model Context Protocol (MCP). It connects generative AI models (Claude, Cursor, custom agents) directly to live edge gateways over standard JSON-RPC 2.0.

---

Architecture: Bridging AI Models and OT Telemetry

Large Language Models (LLMs) cannot safely interact directly with raw Modbus PLC registers or MQTT wildcards. rekuiper-mcp provides a structured, sandboxed boundary:

How It Works

1. Model Context Protocol (JSON-RPC): The AI assistant communicates with rekuiper-mcp via stdio transport.
2. Schema & Stream Inspection: The assistant queries available edge streams, schema types, and active rules.
3. Offline SQL Validation (ruletest): The assistant tests generated SQL against real or simulated message payloads before deploying.
4. Controlled Gateway Deployment: Once verified, the assistant deploys the rule through rekuiper's REST API.

---

Client Setup and Configuration

Add rekuiper-mcp to your MCP client configuration file:

1. Antigravity IDE & Cursor (.cursor/mcp.json or mcp_config.json)

{
  "mcpServers": {
    "rekuiper": {
      "command": "rekuiper-mcp",
      "args": [
        "--server-url",
        "http://127.0.0.1:9081"
      ],
      "env": {
        "RUST_LOG": "info"
      }
    }
  }
}

2. Claude Desktop (claude_desktop_config.json)

• Path on macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Path on Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "rekuiper": {
      "command": "rekuiper-mcp",
      "args": [
        "--server-url",
        "http://127.0.0.1:9081"
      ]
    }
  }
}

---

Built-in MCP Tools for Industrial Agents

rekuiper-mcp exposes a complete toolset for autonomous agents:

Tool Name	Parameters	Purpose
list_streams	None	Discovers all active input telemetry streams on the gateway.
get_stream	name	Returns schema fields, data types, and transport configuration (MQTT, Modbus, OPC-UA).
validate_sql	sql	Checks SQL syntax and type safety offline using the rekuiper query parser.
test_sql_expression	sql, mock_data	Simulates rule execution against synthetic or captured payloads in memory.
list_rules	None	Lists running, paused, and stopped stream processing rules.
create_rule	id, sql, actions	Deploys a verified streaming pipeline to the gateway.
get_rule_status	name	Inspects messages in, messages out, processing latency, and error counters.
list_fleet_devices	None	Inspects connected edge gateways, versions, and heartbeat telemetry.
toggle_wire_tap	rule_id, node_id, enabled	Enables non-intrusive tuple inspection on any DAG execution stage.

---

Real-World Conversational Example

Here is a real conversation between an industrial automation engineer and an AI assistant connected via rekuiper-mcp:

Engineer:
"Inspect our live boiler telemetry stream. If temperature exceeds 120°C for three consecutive readings, write an alert rule that publishes to alerts/boiler_overheat and logs an emergency record."

AI Assistant (via rekuiper-mcp):

1. Invokes describe_stream("boiler_telemetry") to check sensor property names.
2. Writes SQL query:

   SELECT deviceId, temperature, meta(topic) AS source_topic
   FROM boiler_telemetry
   WHERE temperature > 120.0
   GROUP BY deviceId, SLIDINGWINDOW(ss, 10)
   HAVING count(*) >= 3

3. Invokes dry_run_rule with 5 mock events to verify that normal readings are ignored and 3 spikes trigger exactly one alert.
4. Invokes create_rule("boiler_overheat_monitor", sql, actions).
5. Reports: "Rule boiler_overheat_monitor is now active on the gateway. Current status: Running, 0 errors."

---

Safety and Guardrails for Industrial Operations

• Zero Direct Process Control: AI models never write directly to raw hardware registers. All actuation passes through strictly validated rekuiper sink definitions.
• Deterministic Verification: Every generated rule must pass validate_sql and dry_run_rule before registration.
• Poison-Pill Protection: Malformed payloads generated during testing route to Dead-Letter Queues without halting existing rules.

---

Next Steps

• Read the Full MCP Specification
• Explore Rule Testing and Debugging
• Review Dead-Letter Queues and Poison-Pill Isolation