mcp_io 0.2.1

Universal I/O Backbone for device integration into the MCP ecosystem. Provides device registry, policy engine, audit trail, streaming, command queue, and session management.

MCP IO

Universal I/O backbone for device integration into the MCP ecosystem. mcp_io ships the protocol-independent core — the 4-Primitive Contract (describe / read / execute / subscribe), policy engine with interlock evaluation, plan/commit two-phase execution, job manager with cooperative cancel, telemetry hook, and shared type system from mcp_bundle.

Concrete protocol implementations live in companion packages:

Package	Protocol	Production transport
mcp_io_websocket	RFC 6455 (text + binary)	dart:io WebSocket
mcp_io_http	HTTP REST + SSE	dart:io HttpClient
mcp_io_mqtt	MQTT v3.1.1 / v5	dart:io Socket / TLS / WS
mcp_io_serial	UART / RS-232 / RS-485 / USB-CDC	libserialport FFI
mcp_io_can	CAN 2.0A/B + CAN-FD	Linux SocketCAN FFI
mcp_io_modbus	Modbus TCP / RTU / ASCII	dart:io Socket + RTU-over-TCP
mcp_io_opcua	OPC UA Binary	dart:io Socket
mcp_io_scpi	SCPI / IEEE 488.2	dart:io Socket

What this core ships

• AdapterBase — base class adapters extend; provides 4-Primitive surface (describe / read / execute / subscribe), connection state, emergency stop hook.
• Models — DeviceDescriptor, ReadSpec / ReadResult / PayloadEnvelope, Command / CommandResult, TopicSpec, all re-exported from mcp_bundle.
• PolicyEngine — 6-stage pipeline (manifest version → capability match → SafetyClass guard → bounds check → interlock → approval).
• Plan / Commit — planExecute(...) returns a plan token; the call is only sent on commitExecute(token).
• JobManager — cooperative-cancel long-running operations surfaced as io.list_jobs / io.get_job / io.cancel_job.
• IoMetrics — pluggable telemetry sink (NoopIoMetrics / InMemoryIoMetrics shipped; production sinks integrate with whichever observability stack the host project uses).
• MCP-tool dispatcher — exposes the 4-Primitive surface as MCP tools so an LLM (or any MCP client) can talk to any registered adapter without protocol-specific glue.

Quick start

import 'package:mcp_io/mcp_io.dart';
import 'package:mcp_bundle/mcp_bundle.dart';

// 1) Build an adapter (one of the protocol packages above).
final adapter = MyHttpAdapter(...);

// 2) Wire it into the runtime.
final runtime = IoRuntime();
runtime.register('sensor-1', adapter);

// 3) Drive it through the 4-Primitive surface.
final result = await runtime.read(
  deviceId: 'sensor-1',
  spec: const ReadSpec(targets: ['/temperature']),
);
for (final item in result.items) {
  print('${item.uri} → ${item.envelope?.payload.value}');
}

// 4) Subscribe for streaming reads.
final stream = runtime.subscribe(
  deviceId: 'sensor-1',
  spec: const TopicSpec(uri: '/temperature', mode: TopicMode.poll),
);
stream.listen((env) => print(env.payload.value));

Support

• Issues: https://github.com/app-appplayer/mcp_io/issues
• Discussions: https://github.com/app-appplayer/mcp_io/discussions

License

MIT — see LICENSE.