Sparkplug B MCP Server

A Model Context Protocol (MCP) server for Sparkplug B — the Eclipse-standard MQTT topic/payload convention for industrial/SCADA data (birth/death certificates, sequenced metrics, protobuf payloads). Works against any MQTT broker that carries Sparkplug B traffic; only the open Sparkplug B specification is used, no vendor-specific extensions.

This server acts as a Sparkplug B Primary Host Application: it subscribes to the entire spBv1.0 namespace, discovers Edge Nodes and Devices from their NBIRTH/DBIRTH messages, tracks their live metrics, and can send them commands (NCMD/DCMD) — the same role a SCADA/historian host application plays in a real deployment.

Built with FastMCP and pysparkplug (itself built on paho-mqtt).

Tools

Why watching is session-based

Like the MQTT wire protocol underneath it, Sparkplug B is a continuous stream, not a request/response API — a single MCP tool call can't "wait forever" for the next birth or data message. start_watch registers an optional filter (group/node/device/message-type) and buffers matching decoded messages in the background (this connector is always subscribed to the whole spBv1.0 namespace, so no extra broker subscription is needed per watch); get_events drains that buffer, optionally waiting a bounded amount of time for new data. Call stop_watch when done.

Supported metric datatypes

INT8/INT16/INT32/INT64, UINT8/UINT16/UINT32/UINT64, FLOAT, DOUBLE, BOOLEAN, STRING, TEXT, UUID, DATETIME, BYTES, FILE, and their *_ARRAY variants. DATASET, TEMPLATE, PROPERTYSET, and PROPERTYSETLIST are part of the Sparkplug B spec but unsupported by the underlying pysparkplug codec, so they're rejected with a clear error rather than silently mishandled.

For send_node_command/send_device_command, value shape depends on datatype: plain int/float/bool/str for scalars, an ISO 8601 string for DATETIME, a base64 string for BYTES/FILE, and a JSON list for *_ARRAY types.

Configuration

Running locally

uv sync
export SPARKPLUG_MQTT_HOST=localhost
export SPARKPLUG_MQTT_PORT=1883
export SPARKPLUG_HOST_ID=my-mcp-host
uv run sparkplugbmcpserver.py

See mcp.json for ready-to-use client configs (stdio and Docker).

Testing against a real broker

This repo's design principles call for testing against a real instance rather than mocks. docker-compose.yml spins up Eclipse Mosquitto (2.x) with an anonymous listener for local testing. Test fixtures use pysparkplug's own EdgeNode/Device classes to act as real simulated PLCs/devices — publishing genuine NBIRTH/DBIRTH/NDATA/DDATA/NDEATH messages and receiving genuine NCMD/DCMD commands over the wire — so coverage is protocol-level, not hand-rolled mocks. Only the "physical" node/device is simulated, since no free real PLC/edge-node hardware exists to test against; the Sparkplug B protocol handling on both sides is real.

docker compose up -d       # start Mosquitto on localhost:1883 (anonymous, local testing only)
uv run pytest -v           # tests/conftest.py also starts/stops the stack automatically
docker compose down        # stop it manually if you started it yourself

License notes

docker-compose.yml and the files under docker/ configure the open-source Eclipse Mosquitto broker (EPL-2.0) for local testing only — they are not part of the MCP server itself.